Since the seminal work of Carlini et al. at CRYPTO 2020, cryptanalytic model extraction has shown neural-networks parameters can be recovered from black-box queries. Existing attacks are largely built around piecewise-linear phenomena. Softmax attention, as the key component of the transformer architecture, presents a different extraction landscape: its nonlinearity is smooth and sequence-dependent, which renders the existing piecewise-linear-based method inapplicable. Recent work has investigated the learnability of a single-head attention model, while in the multi-head case, the parameters of the multi-head attention layer cannot be uniquely identified from value queries alone. In this paper, we propose the first attack against multi-head attention models. We formalize the extractable representative of multi-head attention and give a polynomial-time algorithm for extracting the parameters of the canonical representative model. We also test our algorithm end to end under finite precision, and successfully extract the parameters of a softmax attention model with token dimension 8 and 6 heads to accuracy $2^{-51}$. Moreover, we overcome the limitation that existing parameter extraction algorithms for one-layer single-head Transformers fail when the ReLU feedforward networks (FFNs) include bias terms. The effectiveness of our approach is demonstrated through model extraction attacks in finite-precision experiments. These results show that softmax normalization itself exposes exploitable algebraic structure, extending cryptanalytic extraction beyond ReLU-centric techniques.

Masking is a well-established software countermeasure against side-channel attacks, yet even algorithmically correct masked implementations can leak on real hardware once the compiler has performed instruction selection, register allocation, and stack allocation. Existing approaches either rely on leakage simulation, which is tied to a specific power model and computationally expensive, or on formal verification of the source program, which does not capture the effects introduced by the subsequent compilation stages. We address this gap from within the compiler and propose a leakage detection pass for the Jasmin language, integrated into its formally verified pipeline and operating on the intermediate representation before register and stack allocation. Rather than simulating power traces, the pass implements a configurable, microarchitecture-oriented leakage model that tracks the contact between shares, secrets, random values, and public values, making the root causes of the detected leakage explicit and enabling the detection of masking-order reductions. We validate the pass on 60 dedicated Jasmin test snippets covering all considered leakage sources and category combinations, and intend it as the foundation for a subsequent compiler stage that automatically removes the detected leakage, thus constituting a first step toward secure compilation.

We prove an algebraic analogue of Pataki-Tural lemma (Pataki-Tural, arXiv:0804.4014, 2008) -- the main tool in analysing the so-called overstretched regime of NTRU. Our result generalizes this lemma from Euclidean lattices to modules over any number field enabling us to look at NTRU as rank-2 module over cyclotomic number fields with a rank-1 dense submodule generated by the NTRU secret key. For Euclidean lattices, this overstretched regime occurs for large moduli $q$ and enables to detect a dense sublattice in NTRU lattices leading to faster NTRU key recovery. We formulate an algebraic version of this event, the so-called Dense Submodule Discovery (DSD) event, and heuristically predict under which conditions this event happens. For that, we formulate an algebraic version of the Geometric Series Assumption -- an heuristic tool that describes the behaviour of algebraic lattice reduction algorithms. We verify this assumption by implementing an algebraic LLL -- an analog of classical LLL lattice reduction that operates on the module level. Our experiments verify the introduced heuristic, enabling us to predict the algebraic DSD event. UPD: minor notation fixes.

ML-KEM is IND-CCA2 secure only because of one check inside decapsulation: the receiver re-encrypts the message it recovered and returns the true shared secret only if the result matches the received ciphertext exactly. This is the Fujisaki–Okamoto (FO) check. wolfSSL implemented it in hand-written SIMD assembly, and on two backends it compared fewer than all of the ciphertext bytes. The x86-64 AVX2 path compared 1536 of 1568 bytes; the ARM64 NEON path compared roughly half. These bugs were documented as a weakening of IND-CCA2 security, in that a tampered ciphertext can slip past the check. We show they are worse than that. The bytes the check skips carry the tail of the decryption noise, and that noise is an exact linear function of the secret key. An attacker who varies those unchecked bytes and watches the decapsulation output reads the noise off one coordinate at a time. Stacking the measurements gives an overdetermined linear system in the secret, which we solve by ordinary least squares with no lattice reduction. The measurement is a plaintext-checking oracle, the same primitive that key-mismatch attacks use. What is new is where it comes from. Reading it off the unchecked v-tail, rather than from chosen sparse-u ciphertexts, means it survives even when u is fully validated, as on AVX2, so the standard "validate all of u" hardening does not close it. The price is queries, 10⁵ to 10⁶ against a few thousand for key-mismatch, so the contribution is reach rather than efficiency. We recover most of the ML-KEM-1024 private key end-to-end against the shipped binaries on both backends: 98.0% of the 2048 secret coefficients at 400 ciphertexts on AVX2, and 98.5% at 600 on NEON, reaching the full key with more ciphertexts (the verified reference model recovers all 2048 at about 1300 ciphertexts). The cost appears to track the geometry of which bytes go unchecked more than their number: NEON leaves about 2.5× more coordinates unchecked than AVX2 yet needs more ciphertexts. We conclude that an incomplete FO comparison is a key-recovery vulnerability, and should be triaged as one.

Zero Knowledge Encryption is a term widely used by vendors of cloud-based password managers. Although it has no strict technical meaning, the term conveys the idea that the server, who stores encrypted password vaults on behalf of users, is unable to learn anything about the contents of those vaults. The security claims made by vendors imply that this should hold even if the server is fully malicious. This threat model is justified in practice by the high sensitivity of vault data, which makes password manager servers an attractive target for breaches (as evidenced by a history of attacks). We examine the extent to which security against a fully malicious server holds true for four leading vendors who make the Zero Knowledge Encryption claim: Bitwarden, LastPass, Dashlane and 1Password. Collectively, they have more than 60 million users and 23% market share. We present 12 distinct attacks against Bitwarden, 7 against LastPass, 6 against Dashlane and 6 against 1Password. The attacks range in severity, from integrity violations of targeted user vaults to the complete compromise of all the vaults associated with an organisation. The majority of the attacks allow recovery of passwords. We have disclosed our findings to the vendors and remediation is underway. Our attacks showcase the importance of considering the malicious server threat model for cloud-based password managers. Despite vendors’ attempts to achieve security in this setting, we uncover several common design anti-patterns and cryptographic misconceptions that resulted in vulnerabilities. We discuss possible mitigations and also reflect more broadly on what can be learned from our analysis by developers of end-to-end encrypted systems.

$\sharp\textsf{Pencil}$ is a domain-extended pseudorandom function by Bhaumik et al, accepted at CRYPTO 2026, claiming that it achieves close to $n$-bit security beyond the birthday bound. It is used as the key-derivation layer of the $\sharp\textsf{Pencil}$-CAU authenticated-encryption mode. We show that $\sharp\textsf{Pencil}$ has a birthday-bound collision attack: its front end $\textsf{Sharp}$ compresses the second half $N_2$ of the input through the $(n-8)$-bit value \[ J(N_2) = \operatorname{msb}_{n-8}\bigl({\mathsf{E}}_{K_1}(N_2 || \texttt{0x00})\bigr), \] after which the entire computation is a deterministic function of $(N_1,J)$. Thus, for any fixed $N_1$, distinct values $N_2,N_2'$ satisfying $J(N_2)=J(N_2')$ produce identical $\sharp\textsf{Pencil}$ outputs. Such collisions occur with probability $1-e^{-1}$ after $2^{(n-7)/2}$ queries, $2^{60.5}$ when $n=128$. This yields a PRF distinguisher with constant advantage, contradicting the security bound of Theorem 4. Since $2^{60.5}$ falls below the birthday bound $2^{n/2}$ that the construction was designed to pass, the beyond birthday-bound property does not hold. Given the derived key $K_1$, an explicit collision can be constructed in approximately $10^3$ inverse-cipher evaluations in expectation. The same collision breaks $\sharp\textsf{Pencil}$-CAU, as two nonce-respecting queries can reuse the key and the nonce of the inner GCM instance, yielding the difference of the two plaintexts. The first version of the paper does not have the defect: it left $J$ untruncated, which makes the map injective and invalidates the collisions above. Reverting to it is the remedy we suggest.

Signal is a secure messaging app offering end-to-end security for pairwise and group communications. It has tens of millions of users, and has heavily influenced the design of other secure messaging apps (including WhatsApp). Signal has been heavily analysed and, as a result, is rightly regarded as setting the "gold standard" for messaging apps by the scientific community. We present two practical attacks that break the integrity properties of Signal in its advertised threat model. Each attack arises from different features of Signal that are poorly documented and have eluded formal security analyses. The first attack, affecting Android and Desktop, arises from Signal's introduction of identities based on usernames (instead of phone numbers) in early 2022. We show that the protocol for resolving identities based on usernames and on phone numbers introduced a vulnerability that allows a malicious server to inject arbitrary messages into one-to-one conversations under specific circumstances. The injection causes a user-visible alert about a change of safety numbers, but if the users compare their safety numbers, they will be correct. The second attack is even more severe. It arises from Signal's Sealed Sender (SSS) feature, designed to allow sender identities to be hidden. We show that a combination of two errors in the SSS implementation in Android allows a malicious server to inject arbitrary messages into both one-to-one and group conversations. The errors relate to missing key checks and the loss of context when cryptographic processing is distributed across multiple software components. The attack is undetectable by users and can be mounted at any time, without any preconditions. As far as we can tell, the vulnerability has been present since the introduction of SSS in 2018. We disclosed both attacks to Signal. The vulnerabilities were promptly acknowledged and patched: the first vulnerability was fixed two days after disclosure, while the second one was patched after eight days. Beyond presenting these devastating attacks on Signal's end-to-end security guarantees, we discuss more broadly what can be learned about the challenges of deploying new security features in complex software projects.

We present the first ciphertext-only distinguishing attack on 5-round AES and a key-recovery attack on 6-round AES for all key sizes under ASCII-encoded English-language plaintext distributions, as well as additional ciphertext-only results under uniform ASCII distributions. Our attacks are enabled by a new analytical framework for estimating truncated differential probabilities in 5-round AES, a problem that remains largely unresolved beyond restricted configurations. Existing approaches, relying on statistical sampling, integral cryptanalysis, or differential distribution tables of super S-boxes, are primarily limited to settings with a single active diagonal in the plaintext and a single passive inverse diagonal in the ciphertext. Our method extends this line of work by providing analytical estimates over a substantially broader range of configurations. We construct this framework by combining precise computations of MDS-level transition probabilities with a systematic enumeration of truncated differential characteristic classes. By organizing characteristics into equivalence classes defined by diagonal propagation patterns, we enable structured aggregation of probability contributions. This approach captures configurations with a single active diagonal in the plaintext and arbitrary passive inverse diagonals in the ciphertext, as well as the complementary setting involving multiple active diagonals in the plaintext and a single passive inverse diagonal in the ciphertext. Our results are validated through independent derivations, consistency checks against prior work, and computer-aided enumeration. More broadly, the framework offers a systematic approach to truncated differential analysis of AES and potentially other AES-like SPN ciphers.

We give a structural cryptanalysis of GRAFHEN, a proposed group-based fully homomorphic encryption scheme whose public rewriting rules hide a group representation used for decryption. Under the designers' isomorphism hypothesis, the compatible representations in a finite target $T$ form one free $\mathrm{Aut}(T)$-orbit. Thus every representative induces the same zero predicate. For a fixed publication, the published key-class expression is the expected trial count of uniform verified sampling rather than a lower bound on reconstruction work. More generally, a homomorphism $\psi$ gives the correct zero predicate exactly when $\ker\psi\cap\Gamma\subseteq\Lambda$, where $\Gamma$ is the ciphertext set of the presented group and $\Lambda$ its zero-ciphertext subgroup. After reconstruction, we give a finite-sample distinguishing bound for independent zero queries and an independent challenge. It requires neither uniform sampling nor generation of the full zero subgroup. For matched query and challenge distributions, the target alone puts the advantage above $1/2$ after $14$ queries at degree $7$ and $26$ at degree $11$. Our attack, HEnbane, combines automorphism-reduced search over complete group elements with algebraic forcing from short public consequences. At the level of a generic algorithmic schema, an uncapped whole-element branching fallback makes reconstruction complete for every finite effectively enumerable target with enumerable element-orbit representatives. The released software instantiates this schema for its supported symmetric-group targets and the fixed target $\mathrm{PSL}_2(343)$. The two-seed force chains used by all supplied instances are a fast path rather than a completeness assumption. It decrypts all five released symmetric challenges, with a slowest ten-run mean of $25.4$ seconds. It also solves the author-supplied $\mathrm{SL}_2(343)$ challenge for the revised parameters by reconstructing in the $\mathrm{PSL}_2(343)$ quotient and recovering the ten central lift signs over $\mathbb F_2$, averaging $37.7$ seconds end to end without exact-order hints.

Selfish mining undermines incentive compatibility of proof-of-work blockchains, letting a miner earn disproportionate rewards at a hashrate lower than the majority threshold. A decade of work has asked whether a strategy is profitable, however, far less is understood about when it becomes profitable. Timing is critical since selfish mining operates at a loss before it turns a profit, typically requiring tens of weeks to break even in the classic case. In this paper, we present a holistic study of the time-to-profitability (TTP) of existing selfish mining strategies structured around four contributions. First, in the single-attacker setting, we characterize TTP across the full strategy space and find that TTP-minimizing and profit-maximizing strategies frequently diverge, making attack horizon a critical metric. In particular, under realistic fee dynamics, the use of incentive transactions to recruit honest-but-rational miners enable incentivized strategies to reach profitability up to $15\times$ faster than classic selfish mining at the same hash rate. Second, we explore TTP for the first time in the multi-attacker setting, showing that the difference in strategies between opposing attackers has a dramatic impact on joint-profitability lag. Third, we generalize intermittent selfish mining by exploring temporal composition over the full strategy space and show that its purported benefits are largely overstated. That is, alternating strategies rarely outperform the best static strategy in terms of either TTP or long-term profits. Finally, and building off our earlier findings, we explore adaptive, state-conditioned strategy selection at the difficulty adjustment period (DAP) level. We compare a general-purpose LLM agent against a fixed decision-tree selector, both implementing the same selection criteria. We find that both selectors reliably identify profit-maximizing strategies from observed network conditions, at a low operating cost, lowering the expertise barrier to exploiting adaptive selfish mining.

Levin and Pedersen proposed at Asiacrypt2025 a new verifiable random function (VRF) based on a CGL-analogue hash function constructed from radical isogenies. Their construction applies the same secret radical-CGL walk to a public starting curve and a message-dependent curve, and uses an R1CS proof relation to show that the two walks use the same secret key. We present a two-stage attack on this construction. The first stage concerns the unspecified representation of the public key. The reported key size indicates that the public curve is stored as a \(j\)-invariant, whereas both the specified radical-CGL computation use two coefficients to represent a curve. By exploiting this form we can produce two different VRF outputs under the same public key and message, breaking the unique provability. Hence, the output of the radical-CGL computation must follow the specification. In the second stage, we exploit these coefficients to recover the VRF secret key. With \(1536\) queries, our implementation recovers the complete \(256\)-bit secret in 30 minutes, thereby breaking residual pseudorandomness. Interestingly, we also observe that the using public key alone without queries can sometimes reveal one or two bits of the secret walk. Besides, we extend Lai's observation to obtain a one-query attack on the group-action-based VRF proposed in the same paper with advantage closed to 1/2. Together, these constitute three attacks on their work.

The Unbalanced Oil and Vinegar (UOV) scheme was proposed by Kipnis et al. in 1999 as a multivariate signature scheme. Owing to its small signature size and its resistance to various attacks over more than two decades, UOV has become one of the leading candidates in multivariate public key cryptography. In 2025, Ran proposed a novel algebraic key recovery attack exploiting the algebraic structure of UOV, which reduced the security of several parameter sets of UOV and its variants submitted to the second round of the NIST PQC standardization process for additional signatures. This attack was improved by Jin et al., and Furue and Ikematsu, forming a line of attacks that has significantly advanced the cryptanalysis of UOV. However, Ran's attack is applicable only when $v<2m$, where $v$ denotes the number of vinegar variables and $m$ the number of public polynomials. In fact, when $v\ge 2m$, an additional kernel element of the ideal generated by the public polynomials appears, preventing the attack from recovering the oil subspace. A similar issue arises in the improvements by Jin et al., and Furue and Ikematsu. In this paper, we propose a method that overcomes this issue, extending the applicability of this line of attacks to the case where such an additional kernel element appears. Applying our method to SNOVA via the lifting technique of Nakamura et al., we show that the claimed security levels of some parameter sets of SNOVA in the second round of NIST PQC standardization process for additional signatures are reduced. In particular, for the parameter set $(v,o,q,l)=(37,17,16,2)$ of NIST security level I, although Ran's attack is not applicable, our method reduces the estimated security to $2^{103}$ gate operations, which matches the complexity of the attack by Bros et al. in 2026.

Designated Verifier Signatures (DVS) are an important variant of digital signatures that ensure only a specified verifier can validate a signature, while preserving non-transferability. With the advent of quantum computing, several quantum DVS schemes have been proposed to achieve quantum security. In this paper, we revisit the quantum DVS protocol of Xin et al. [Quantum Information Processing, 2022] and provide a structural cryptanalysis of its design. We show that the scheme admits an existential forgery under a chosen-message attack: given a valid quantum signature on one message, an adversary can efficiently transform it into a valid signature on another message without knowledge of the signer’s private key. To address this weakness, we propose a minimal countermeasure based on QKD-derived keys and quantum one-time pad encryption.

The VOX signature scheme is a multivariate signature scheme which was submitted to the Round 1 Additional Digital Signature Schemes NIST process. VOX relies on the Hat Plus perturbation and the Quotient-Ring transform (QR). We formalize a dimension criterion enabling the direct attack to be used as a key recovery attack against UOV schemes. This enables a practical cryptanalysis of the Round 1 VOX parameters. Next, we show that some of the alternative parameters proposed for VOX after attacks on the Round 1 submission are still vulnerable. More precisely, these parameters were chosen to defeat an attack of Furue and Ikematsu in the field extension defined by the QR parameter. We observe that one may use a smaller field extension of any degree dividing the QR parameter, in which case the attacks apply again. These attacks are relevant for a subset of the parameter sets proposed for VOX: I, Ic, III, IIIa, V, Vb. In particular, we apply the subfield framework to our previous dimension criterion for the direct attack. We estimate the cost of our attack on these parameter sets and find costs of at most 2^67 gates, and significantly lower in most cases. In practice on a commercial laptop, our attack requires 0.3s, 1.35s, 0.56s for parameter sets I, III, V for VOX, and 56.7s, 6.11s for the alternative parameter sets IIIa, Vb. Our analysis also improves the cryptanalysis of some alternative parameters proposed by Guo and Ding, and of the ``Minus'' variant of VOX.

At CRYPTO 2019, A. Gohr introduced Neural Differential Cryptanalysis and used deep learning to improve the state of the art in cryptanalysis of 11-round SPECK32. As of February 2025, Gohr’s article has been cited 238 times on Google Scholar. The variety of targeted cryptographic primitives, techniques, settings, and evaluation methodologies that appear in these follow-up works provides a basis for a careful survey, which we provide in this paper. More specifically, we propose a taxonomy of these 238 publications and systematically review the 71 papers focusing on neural differential distinguishers, pointing out promising directions and recent advances in explainability and key-recovery attacks. We then highlight future challenges in the field, particularly the need for improved comparability of neural distinguishers and advancements in scaling. This holistic survey helps researchers and engineers to identify the leading neural differential attacks, compare their performance, and highlight the outstanding open problems in AI-assisted cryptanalysis.

We present the first passive side-channel distinguisher on LESSv2.0, a second-round candidate in NIST’s call for additional post-quantum digital signature schemes. We target the Gaussian elimination at the core of LESS and and present a method to exploit algorithmic leakage arising from the manipulation of sparse versus dense columns. We show that this leakage, while trivially available in non-constant-time implementations, also persists in constant-time implementations and can be exploited using a distinguisher. Concretely, the proposed attack relies only on distinguishing zero-valued computations from random ones that are repeatedly evaluated during the computation, leading to a large attack surface and making the attack robust to noise. Through simulation, we experimentally show that the required number of observed signatures lies between 300 and 2357 depending on the parameter set. This relatively large number is due to the targeted information being inherently noisy, leading to a correlation-based key recovery attack, even with noiseless leakage. Furthermore, we discuss three common countermeasures: first-order masking, shuffling and blinding. Finally, we validate our approach on implementations with and without masking by showing the presence of leakage on a physical target.

Witness encryption (WE) allows one party to encrypt a message under an arbitrary satisfiable circuit, so that anyone holding a satisfying input can decrypt. Efficient WE enables numerous modern applications, such as identity-based and attribute-based encryption. Recent candidates for efficient WE base their security on rank properties of structured ciphertext matrices, which encode the validity of a given witness. This shrinks ciphertext sizes considerably compared to previous constructions, but rests on heuristic arguments rather than security reductions. We describe two attacks against two such constructions, namely the affine determinant program (ADP) construction from 2020 and its arithmetic extension, the AADP, from 2026. The first attack observes that for sparse circuits, the natural regime for both schemes, commutators formed from the public ciphertext matrices have unexpectedly low rank. Elementary linear algebra on these matrices then recovers the encrypted message directly from the public ciphertext, without knowledge of any witness, and hence breaks the security of both schemes. The second attack linearizes the nearly-skew-symmetric (NSS) variant of the ADP construction, recovering the encryption randomness and the message. To our knowledge, ours are the first attacks against these WE candidates, and we verify both in practice.

HAWK is a lattice signature scheme that is currently a third-round candidate in NIST's post-quantum signature competition. We give an unconditional, deterministic polynomial-time reduction from HAWK-$n$ key recovery over $K_n=\mathbb{Q}(\zeta_{2^\ell})$ to $\mathrm{poly}(n)$ calls to an exact Shortest Vector Problem (SVP) oracle in dimension $n/2+1$, where $n=2^{\ell-1}$ is the ring degree. The reduction uses a nontrivial automorphism of the key lattice, supplied by the Galois involution $\tau:\zeta\mapsto-\zeta$ and recoverable as a shortest vector of a public rank-$n$ lattice isometric, up to scaling, to $\mathbb{Z}^{n/2+1}\oplus\sqrt{2}\,\mathbb{Z}^{n/2-1}$. Ducas's block reduction on this near-hypercubic class finds the automorphism, and the descent of van Gent and Pulles recovers the key from it. In the gate-count model, the attack lowers the key-recovery cost of HAWK-512 from $2^{150}$ to $2^{108}$ and of HAWK-1024 from $2^{288}$ to $2^{182}$. We demonstrate this with a practical implementation that recovers a HAWK-256 secret key end-to-end in a few hours on a single server. The construction does not transfer to Falcon. Conductors $m\in\{p^k,2p^k\}$ ($p$ an odd prime), i.e.\ the $m>4$ with cyclic $(\mathbb{Z}/m)^\times$, evade the attack.

Pseudorandom error-correcting codes (PRCs), a novel cryptographic primitive recently proposed at CRYPTO 2024, are primarily applied in undetectable watermarking schemes for large generative models. However, the security of PRCs has not yet been systematically analyzed. To fill this gap, we present the first cryptanalysis of PRCs. Specifically, focusing on LDPC-PRC, the only known practical instantiation of PRCs, we propose three novel attacks that challenge its undetectability and robustness. To rigorously demonstrate the practical threat, we analyze the concrete attack complexity under realistic parameters and validate the attack effectiveness on both real-world large language models and generative image models, including DeepSeek and Stable Diffusion. Our analysis shows that the claimed security guarantees of LDPC-PRC are undermined across all practically feasible regimes. For example, our attacks can detect the presence of a watermark with overwhelming probability at a cost of $2^{22}$ operations. Beyond attacks, we further propose three defenses: parameter recommendation, implementation suggestion, and a revised key generation function. However, PRC-based watermarking schemes still fail to achieve 128-bit security due to inherent constraints of large generative models, such as the maximum output length of large language models. Overall, our work clarifies the concrete security limits of PRCs in real-world watermarking applications.

Let $E=\mathbb Q(\zeta_m)$ be a power-of-two cyclotomic field, with maximal totally real subfield $K=\mathbb Q(\zeta_m+\zeta_m^{-1})$. In previous work, Chevignard et al. (Eurocrypt'25) gave a reduction from module-LIP for rank-two module lattices over $\mathcal O_E$ to the norm-reduced Principal Ideal Problem (nrdPIP) in a quaternion algebra. We derive two consequences of this reduction. First, we obtain a polynomial time reduction from rank-$2$ module-LIP over $\mathcal O_E$ to rank-$3$ module-LIP over $\mathcal O_K$. Note that rank-$2$ modules over $\mathcal O_E$ are naturally seen as rank-$4$ modules over $\mathcal O_K$, so this is indeed an improvement. Our second result is specific to the module $\mathcal O_E^2$, underlying the Hawk signature scheme. In this setting, we also obtain a reduction to a rank-$3$ module-LIP instance over $\mathcal O_K$, but we can additionally show that these modules have a simple geometric shape: they are isomorphic to $\mathbb Z^{m/2+1} \perp \sqrt{2}\, \mathbb Z^{m/4-1}$. We then adapt Ducas' result (ePrint'23) to this setting. Putting everything together we obtain an algorithm breaking Hawk's key recovery by making polynomially many exact-SVP calls in lattices of dimension at most $3m/8+1$. This improves upon the previous analysis from Ducas (ePrint'23) which required exact-SVP calls in lattices of dimension at most $m/2+1$.

Elliptic Curve Cryptography (ECC) underpins modern public-key infrastructure, relying on the computational hardness of the Elliptic Curve Discrete Logarithm Problem (ECDLP). While monolithic quantum architectures running Shor's algorithm threaten ECC, their physical realization is bottlenecked by massive logical qubit demands for modular inversion. Distributed Quantum Computing (DQC) offers a scalable pathway by interconnecting smaller, cooperative Quantum Processing Units (QPUs). In this paper, we adapt two state-of-the-art distributed quantum frameworks, namely a zero-quantum-communication paradigm and a sequential teleportation-based protocol, to the elliptic curve setting. We provide the first rigorous resource estimation of distributed ECDLP solvers. By incorporating the compact, register-sharing Extended Euclidean Algorithm (EEA) formulation, we quantify the physical design trade-offs for a cryptographically secure 256-bit elliptic curve on fault-tolerant architectures. Specifically, the zero-quantum-communication variant operates with 1094 to 1154 logical qubits per node, whereas the teleportation-based variant requires as few as 856 to 1098 logical qubits per node. Furthermore, optimizing the arithmetic window size ($\omega = 16$) over $k = 22$ nodes reduces the single-node Toffoli gate count to $2^{26.84}$, which represents an approximate $14\times$ reduction compared to the $2^{30.63}$ baseline of monolithic schemes. Our results delineate the architectural boundaries between inter-QPU communication, classical coordination, and single-chip hardware floors, mapping out an efficient design space for distributed quantum cryptanalysis.

This note presents a new unconditional attack on the supersingular isogeny problem, with time and memory complexity $p^{2/5+o(1)}$. It builds on the approach by Eisenträger-Hallgren-Leonardi-Morrison-Park (2020) and Fuselier-Iezzi-Kozek-Morrison-Namoijam (2025), and is related to the recent heuristic attack with complexity $p^{1/3+o(1)}$ by Wesolowski (ePrint 2026/1486): all of these search for a separable isogeny from a curve to its Galois conjugate to form a non-scalar endomorphism. Our attack is based on highly theoretical multivariate multipoint evaluation algorithms from Kedlaya-Umans (2008, 2011), Bhargava-Ghosh-Guo-Kumar-Umans (2022), and Ghosh-Harsha-Herdade-Kumar-Saptharishi (2023), and therefore does not threaten isogeny cryptosystems in practice; it is of theoretical interest.

The security of a recent MPC-in-the-Head signature scheme introduced at Crypto 2024 by Huth and Joux required the introduction of a new ad hoc hardness assumption called the Subfield Bilinear Collision (SBC) problem. By relying on the VOLE-in-the-Head framework, the authors further improved the performance of this scheme at Asiacrypt 2025, resulting in a very compact construction. In this paper, we improve the original cryptanalysis of SBC in several ways. First, we describe a link between the SBC problem and the decoding problem in the rank metric for codes linear over an extension field $\mathbb{F}_{q^m}$, strengthening its theoretical hardness and expanding the range of attacks on the SBC problem. Second, we analyze Gröbner basis algorithms applied to the bilinear modeling of SBC proposed by Huth and Joux and formulate conjectures on the behavior of this system. Finally, we describe another algebraic modeling of SBC obtained by using the Plücker relations between the maximal minors of a matrix. While we do not threaten the parameters of the proposed signature schemes relying on SBC, our work opens the door to a more accurate analysis. In particular, we were not able to analyze the inclusion of the field equations of $\mathbb{F}_{q}$ in the Gröbner basis algorithm, which is especially relevant since $q=2$ in these parameters. We leave this task for future work.

Hamming Quasi-Cyclic (HQC) has been selected by NIST for standardization in the post-quantum landscape. As deployment approaches, implementation security becomes as critical as mathematical hardness. In this work, we demonstrate that source-level constant-time coding is not a standalone guarantee: the compiled binary must inherently preserve this behavior. We identify a severe compiler-induced vulnerability within the official AVX2-optimized implementation of HQC, despite its claims of being constant-time. Although the C source code relies on secure, mask-based conditional selection, certain compiler optimizations rewrite this logic systematically. This transformation silently introduces secret-dependent control flow into the inner Reed-Muller decoding process, resulting in secret-dependent cache access patterns. Exploiting this vulnerability, we mount, to the best of our knowledge, the first cache-timing Full-Decryption-style oracle attack against a post-quantum cryptosystem. Using Flush+Reload on shared libraries, an unprivileged co-located adversary can extract fine-grained predicates of the decoder's internal state. To achieve full key recovery, we develop a novel, reliability-aware Soft Information Set Decoding (Soft-ISD) post-processing framework. Leveraging a GPU-accelerated meet-in-the-middle strategy optimized for heterogeneous platforms (including Apple Silicon), we demonstrate end-to-end secret key recovery for hqc-1 with less than 10 seconds of online trace collection.

UOV and its variants have been submitted to NIST's PQC Standardization Project. And security analysis against UOV families (i.e., UOV and its variants) are typically categorized into key-recovery attacks and forgery attacks, with the XL algorithm serving as one of the most significant methods for mounting key-recovery attacks. Recently, Lars Ran introduced a novel key-recovery attack against UOV families that could be seen as an XL attack using exterior algebra; nevertheless, this new XL attack is applicable only when the underlying (finite) field of UOV instances is of characteristic $2$. In this work, we address this limitation by generalizing Lars Ran's approach, and our contributions are three-fold. First, we generalize the notion of exterior algebra and propose the notion of reduced symmetric algebra over any field, which is essentially symmetric algebra when the characteristic $p$ of the underlying field is $0$ and is exterior algebra when $p=2$. Moreover, based upon the reduced symmetric algebra, we propose a new XL attack against all UOV families: as a generalization of Lars Ran's attack, our XL attack also works for those UOV instances with odd characteristic; for instance, it works for those $12$ recommended QR-UOV instances submitted to NIST's PQC Standardization Project, although its performance does not outperform those of the existing key-recovery attacks. Finally, analysis on our XL attack reveals the connection between the characteristic $p$ and the concrete hardness of UOV instances.

We present new sample-runtime tradeoffs for the decisional sparse Learning With Errors (LWE) and sparse Learning Parity with Noise (LPN) problems over $\mathbb{Z}_q$, specifically in regimes where the secret vector is constrained by a small $l_{\infty}$ norm. While small-secret constraints are useful for the practical efficiency of lattice-based cryptography—such as homomorphic encryption and zero-knowledge proofs—the extent to which an adversary can exploit these bounds when the coefficient matrix is sparse is an open question. We address this by reducing the distinguishing task to a relaxed variant of the Short Integer Solution (SIS) problem, where the strict $A^\text{T} \mathbf{c} = 0$ requirement is replaced with an $l_1$-norm bound on $A^\text{T} \mathbf{c}$. To solve this relaxed SIS problem, we design an algorithm that samples distinct, non-trivial walks on a Kikuchi graph having close end points. For LWE, this approach directly separates planted from random instances. For LPN, where the noise is uniformly distributed over non-zero elements, the proof is more involved. We first derive a different reduction from LPN to (relaxed) SIS and then extend the anti-concentration framework given by Gupta, He, O'Donnell, and Singer (SODA 2026).

PRISM (PKC25) is a hash-and-sign signature scheme whose security relies on the hardness of computing large-prime-degree isogenies originating from a curve of unknown endomorphism ring. In PRISM, the degree of such isogenies is obtained by hashing messages onto a set of large odd integers that pass a primality test. In this work, we investigate the impact of the choice of primality test on the security of PRISM. We first show that when a weak primality test is used, the assumption underlying the security proof in the standard model does not hold. We then extend our analysis to the assumption used in the security proof in the (quantum) random oracle model. In this setting, we argue that the Miller-Rabin test suffices and estimate the minimal number of iterations required for PRISM to achieve the desired security level, thus minimising signing costs.

At ASIACRYPT 2021, Baksi et al. introduced DEFAULT, a block cipher designed to algorithmically resist Differential Fault Attack (DFA), claiming 64-bit DFA security regardless of the number of injected faults. At EUROCRYPT 2022, Nageler et al. demonstrated that DEFAULT's claimed DFA resistance can be broken by applying an information-combining technique. More recently, at ASIACRYPT 2024, Jana et al. improved DFA by searching for differential trails with a single solution. They showed that, for DEFAULT with a simple key schedule, injecting five faults at the fifth-to-last round reduces the key space to one, and for BAKSHEESH, injecting twelve faults at the third-to-last round achieves the same result. In this paper, we propose a new DFA framework that utilizes a Mixed-Integer Linear Programming (MILP) solver. This framework makes it possible to attack deeper rounds than previously achieved, reducing the number of fault injections required for key recovery. Furthermore, we present a method to determine the most efficient fault injection bit positions by systematically analyzing the input differences from all possible single bit-flip faults, thereby further reducing the required number of faults. This systematic analysis has the significant advantage of allowing us to theoretically calculate the required number of faults. Applying our framework, for DEFAULT, injecting three faults at the sixth-to-last round and two faults at the seventh- and eighth-to-last rounds reduces the key space to one.

Neural differential distinguishers have become an active research direction in​ symmetric-key cryptanalysis since the introduction of deep-learning-based attacks on​ round-reduced SPECK. Early neural distinguishers typically used a single ciphertext pair​ or ciphertext difference as input. Recent studies, however, show that richer input​ representations can substantially affect the information available to the classifier, the data​ cost of each labeled sample, and the relevance of the distinguisher to practical attacks.​ Examples include multi-pair, multi-difference, matrix-style, multi-round,​ structured-encoding, and score-aggregation based inputs.​ This paper provides a taxonomy and survey of rich input representations in neural​ differential cryptanalysis. We introduce a representation-centric framework that describes​ an input representation by its difference set, number of observations per sample, sharing​ structure, encoding function, and ciphertext cost. Using this framework, we organize​ existing works into representation families and compare their motivations, benefits, and​ limitations. We also argue that representation-rich distinguishers require cost-aware​ evaluation: fixed-sample comparisons and fixed-ciphertext comparisons answer different​ questions and may lead to different conclusions. Finally, we identify open problems related​ to automated representation search, theoretical explanation of representation gain,​ cipher-family transferability, interpretability, reproducibility, and key-recovery integration.​ The survey highlights that rich input representations should be treated as first-class​ cryptanalytic design choices rather than secondary implementation details.

In AfricaCrypt 2025, Seck et al. proposed a new generalized Wiener-type attack on an RSA-like cryptosystem proposed by Cotan and Teseleanu (NordSec 2023). In their attack, they studied the generalized key equation $eu - (p^4 - 1)(q^4 - 1)v = w$ and showed that a private exponent $d$ which is too large or too small can be recovered in polynomial time. Another RSA variant based on cubic Pell curves with key equation $ed - (p - 1)^2(q - 1)^2 k = 1$, was examined by Rahmani and Nitaj in AfricaCrypt 2025. Note that these two attacks are valid for a balanced modulus $N = pq$ ($q < p < 2 q$). In this paper, we extend these two attacks by showing that for a modulus $N=pq$ product of arbitrary primes $p$, $q$, one can efficiently factor $N$ by studying the two key equations $ex - (p^4 - 1)(q^4 - 1)y = \omega$ and $ex - (p - 1)^2(q - 1)^2 y = \omega$ under certain conditions on $x,y$ and $\omega$. Our new attacks are based on Coppersmith method and continued fractions.

Simon's algorithm can detect hidden XOR periods in functions derived from symmetric ciphers. Finding such functions becomes difficult when nonlinear layers and diffusion spread the relevant expressions across many branches, so recent work has used symbolic search to automate the construction. We refine the algebraic SMT model of Liu et al. in two ways. Prefix realization checks whether a symbolic starting state can be reached through preceding rounds and records the round-key nibbles needed to produce it. DDT Filtering restricts a local S-box input to a DDT bucket so that the symbolic path can cross an additional nonlinear layer. The latter condition is key-dependent: the target period need not lie in the translation space of the selected bucket, and our results state this condition explicitly. We report the maximum round counts found for GFS-2F, GFS-4F, Skipjack-B, LBlock, TWINE, CRAFT, and SKINNY, with Liu et al.'s automated model as the main comparison. We also combine selected witnesses with partial round-key guesses in the Grover–meet–Simon setting, yielding reduced-round key-recovery candidates below the corresponding comparison budgets.

Hybrid fully homomorphic encryption (FHE) inference improves the practicality of private inference by letting the server evaluate linear layers homomorphically while the client decrypts and applies nonlinearities. Recent schemes attempt to protect model confidentiality by returning noisy, output-permuted responses and appealing to shuffle-model differential privacy (DP). We show that this protection fails in the correctness regime required by hybrid FHE systems. For a $d$-input linear layer, $d+1$ admissible queries suffice for exact recovery of a permutation-invariant layer summary, hence for perfect model distinguishability. We further show that input DP is orthogonal to model confidentiality and that the local-DP premise required for shuffle amplification cannot hold under correctness-bounded noise. We recover all linear layers of a SAFHIRE-style ResNet-20 end-to-end from TFHE transcripts with zero error, using $d+1$ queries per layer for a total of $5{,}712$ direct queries. Under the same query model, we also confirm exact per-layer recovery on pretrained ImageNet-scale CNNs and ViT-B/16. The leaked spectra enable fingerprinting, lineage attribution, and improved logit-based extraction, while suppressing them destroys inference utility.

Despite decades of research in electromagnetic (EM) side-channel analysis (SCA), practical attacks still require manual effort and domain expertise to identify informative probe locations on target devices. Existing approaches rely heavily on exhaustive grid scanning or handcrafted alignment, limiting attack scalability and realism. In this work, we present the first automated and adaptive EM SCA framework that uses particle swarm optimization (PSO) to navigate the probe. Particles are guided by mutual information (MI) leakage maps, enabling efficient recovery of secret-dependent emissions. We introduce a novel application of the Nyström approximation to accelerate MI estimation across EM trace windows, allowing real-time swarm guidance without full kernel computations. Unlike prior work, our method requires no leakage templates, manual tuning, or alignment assistance, enabling automated attacks with minimal assumptions. We validate our framework on both microcontroller and FPGA platforms running AES-128. PSO-guided scanning identifies high-leakage points faster than grid search and reduces the number of traces required for successful CPA-based key recovery by up to a factor of 16, i.e., saving tens of thousands of traces.

The impossible boomerang (IB) attack is one of the most powerful cryptanalytic techniques for block ciphers, with notable successes against AES. However, constructing IB distinguishers—the core component of IB attack—remains a fundamental challenge: existing methods are restricted to specific cipher structures, cannot detect contradictions spanning more than two rounds, and lack a unified framework across attack settings. To solve these issues, this paper introduces three hierarchically complementary methods for constructing IB distinguishers: (1) a GEBCT-based method breaking the 2-round contradiction barrier; (2) a state-based method proven exactly equivalent to the definition of IB distinguishers (the most precise available); and (3) a structural method for fast security evaluation. We prove their inclusion relations, unify all prior methods, and connect the state-based method to the Quasidifferential Framework within IB distinguishers. Furthermore, we develop a unified SAT/SMT-based framework that automatically searches distinguishers under single-key, two-related-key, and four-related-key settings. We evaluate on seven representative block ciphers covering almost major design paradigms: We present the first IB distinguishers for ARX cipher SPECK (ISO/IEC standard) and CHAM, bit-permutation cipher GIFT-64 (CHES'17, an example with 5-round contradictions), and key-dependent Sbox cipher PRINTcipher48 (CHES'20). We obtain the first 5-round related-key IB distinguishers for AES-128 (NIST's standard with nonlinear key schedule, first automatically searched), and 1-round improvements over prior IB distinguishers for tweakable cipher SKINNY-64/192 (CRYPTO'16) and SKINNYee (CRYPTO'22). All IB distinguishers outperform the best impossible differentials by 1--4 rounds, including a PRINTcipher48 result that disproves the long-standing claim about single-key IB distinguisher limitations. Finally, we present a 31-round related-key IB attack on SKINNYee, the best result to date.

The Generalized Feistel Network (GFN) underpins widely standardized block ciphers, yet its resistance to full plaintext recovery remains largely unexplored. This paper extends the full-plaintext attack framework for standard Feistel ciphers to multi-branch scenarios, mainly makes the following $3$ research contributions: (1) Developed classic full plaintext recovery attacks on $d$-round Type-I GFN ($d\ge 3$) under CPA with $d+1$ encryption queries and $2d$-round Type-I GFN under CCA with $d$ decryption queries. Query complexity depends on d, increasing branches can enhance anti-interference capability. (2) Developed classic full plaintext recovery attack on 2-round Type-II GFN ($d\ge 4$) under CPA with $3$ encryption queries and 3-round Type-II GFN under CCA with $1$ encryption plus $2$ decryption queries. Query complexity is independent of $d$, increasing branches does not improve resistance. (3) Finded that all attacks treat round functions as black boxes, confirming that the weakness resides in the GFN topology rather than specific round-function designs. Strengthening S-boxes or diffusion matrices cannot mitigate it, only increasing rounds beyond the security threshold provides effective defense.

GRAFHEN is a group-based homomorphic-encryption proposal whose public key is a rewriting system and whose secret key is a permutation representation. We present Maverick, an equivalent-key recovery attack. Maverick breaks every released GRAFHEN challenge, including the recommended two-copy $S_{11}$ instance: it reconstructs an equivalent key from public rules and correctly decrypts all $20{,}000$ supplied labelled ciphertexts. On $12$ independently generated recommended-parameter keys, the median end-to-end time is $552$ s and the median peak memory use is $11.28$ GB on an Apple M3 Pro. The attack converts selected public rewrite rules into group relators, reconstructs a regular action by Todd-Coxeter enumeration, recognizes the resulting permutation representations, and aligns the two sides through the public mixed relations. Public labelled encryptions then calibrate an equivalent decryptor. We prove a proof-carrying version of this procedure: a target-order table with a replayable trace certifies the recovered regular action. The reported $S_{11}$ experiments use target-order closure checks, complete-corpus verification, and decryptor validation. We also give an output-sensitive analysis and state the hypotheses needed to extrapolate beyond the measured instances. Following an independent key-recovery attack, GRAFHEN proposed in July 2026 to replace $S_{11}$ by $\mathrm{PSL}_2(343)$. We adapt Maverick to this setting and demonstrate complete public-rule recovery, alignment, certification, and calibration on generated $\mathrm{PSL}_2(169)$ instances. The two-side $q=169$ run decrypts all $5{,}000$ held-out ciphertexts in a $103$~s critical path using $0.93$ GiB peak memory; a $k=2$ admissibility-filtered corpus also succeeds. Under GRAFHEN's stated worst-case estimate for Dumezy's degree-based search, this degree $170$, $d=5$ setting already has cost $O(2^{850})$, well beyond the intended reach of that attack. At the target group, the post-enumeration recovery path from a generated regular action takes $35.0$ s and $1.92$ GiB peak memory. These results suggest that the proposed platform-group change is insufficient to rule out Maverick; recovery from an admissibility-filtered corpus at $q=343$ remains to be measured because the pre-filter KeyGen enumeration exceeded $32$ GiB of memory before it could produce a corpus.

Almost fifty years after its introduction, the McEliece cryptosystem occupies an unusual place in the post-quantum landscape. Its public keys are far larger than those of most competing schemes, its original parameters no longer provide adequate security, and several compact variants proposed to reduce key size have subsequently been broken. Nevertheless, the binary Goppa-code foundation retained in Classic McEliece continues to resist known practical attacks for the selected Classic McEliece parameter sets. This survey asks why McEliece has remained relevant despite these limitations. We trace its development from the original 1978 encryption scheme to the modern Classic McEliece key-encapsulation mechanism and organize nearly five decades of cryptanalysis into generic decoding, structural recovery, attacks on compact variants, protocol-level attacks, implementation leakage, and quantum speedups. We emphasize several distinctions that are often blurred in discussions of the scheme: breaking an obsolete parameter set is not the same as recovering the hidden Goppa structure; distinguishing a public code does not necessarily lead to practical key recovery; and compromising a modified or structured variant does not automatically compromise Classic McEliece. This history does not support either of two simple narratives: that McEliece has remained unchanged or that it has simply been broken. Its longevity reflects a conservative mathematical foundation that has survived repeated reassessment, together with parameters, security models, and implementations that have evolved in response to new attacks. We close by outlining the main questions that will shape its future: whether public-key and key-distribution costs can be reduced without exposing exploitable structure, how far modern algebraic cryptanalysis can be extended, how classical and quantum security estimates should be refined, and how secure implementations can be integrated into practical systems.

Standard Tseitin encodings of the SHA-256 nonlinear functions Ch and Maj can hide conditioned differential projections from Boolean Constraint Propagation (BCP). We materialize them as short, semantically redundant CNF clauses. A cofactor theorem characterizes all controlled differential linear forms; its implemented unit-vector specialization returns exactly all minimum-control projections, yielding four Ch and twelve Maj clauses per bit. The clauses preserve models, introduce no variables, and strictly strengthen BCP on an explicit gate fragment. We claim neither propagation completeness nor a general affine compiler. We evaluate mechanism separately from performance and distinguish the production bundle from the proposed layer. Frozen studies show a fixed-formula bundle benefit, but the matched clause isolation fails its effect gate and an unrestricted control is inconclusive. We therefore show neither a solver-independent speedup nor a new cryptanalytic attack. The restricted weight-15 C15 census passes the CaDiCaL rule but not the Kissat rule. In the stratified C16 extension, the frozen decisions are too censored for CaDiCaL and too censored for Kissat; these solver-stratified labels are not pooled. A pinned three-solver replay validates larger-bundle execution but cannot attribute performance to the clauses. A variable-preserving probe exposes all 32 tested implications only after augmentation.

For cryptosystems whose security relies on the hardness of decoding in the sublinear regime, the best known attacks are based on Information Set Decoding (ISD). In this regime, which is particularly relevant to HQC and Pseudorandom Correlation Generators (PCG), the cost of Gaussian elimination is no longer negligible and significantly affects the overall attack complexity. In this work, we revisit the Reduce-and-Prange technique of Kim and Lee, which reduces the cost of Gaussian elimination by reusing partial pivots. We refine its complexity analysis using branching-process techniques, thereby obtaining a more accurate assessment of its performance. We then extend partial pivot reuse to Stern's algorithm and introduce MAGE-Stern, a multilevel amortized Gaussian elimination variant of Stern's algorithm. Under a consistent logic-gate cost model, MAGE-Stern improves upon the best previously known attack against HQC by approximately 3 bits in time complexity, while reducing the memory complexity by about 12 bits. In particular, we estimate the security of the standardized HQC Category I parameter set at approximately 140 bits, about 3 bits below its NIST security target. We further combine multilevel amortized Gaussian elimination with the projective decoding framework of Carrier, Hatey, and Tillich, and investigate its application to regular decoding. Applied to the reference Pseudorandom Correlation Generator (PCG) parameter sets of Boyle, Couteau, Gilboa, and Ishai, the resulting algorithms improve upon the best previously known attacks by up to 6 bits across a broad range of practical parameters.

We prove that under a plausible heuristic assumption (on the smoothness of certain random integers), the supersingular isogeny problem can be solved in time and memory $p^{1/3 + o(1)}$. This improves upon the previous best complexity of $p^{1/2} \cdot(\log p)^{O(1)}$. This problem is arguably the central hard problem underlying isogeny-based cryptography, and the cost of its resolution is a major (and often the only) factor in the choice of secure parameters. The impact on concrete parameter sets remains to be clarified, as the asymptotic advantage of the new algorithm is mitigated by a superpolynomial overhead hiding in the $o(1)$ exponent, and by its high memory requirement.

We make progress on the question of constructing explicit evaluation places for leakage-resilient Shamir's secret sharing, over composite order fields. Previously, Maji et al. (EUROCRYPT 2024) showed that random evaluation places yield Shamir's secret sharing over the composite order field $\mathbb{F}_{p^d}$ that is statistically secure against physical-bit leakage. Later, Nguyen (EUROCRYPT 2025) established a "dichotomy" that linear code-based secret-sharing scheme over the field $\mathbb{F}_{p^d}$ is either statistically secure or completely insecure against such leakage. Building upon Nguyen's dichotomy, we present a partial derandomization of evaluation places, improving upon the Maji et al. result for a restricted regime of parameters. We replace the random choice of $n$ independent evaluation places by the iterates $x_j = \Phi^j(x_0)$ of a simple fixed rational function $\Phi$, where the initial point $x_0 \in \mathbb{F}_{p^d}^*$ is randomly chosen. The randomness in the evaluation places thus drops from $nd \log p$ bits to $d\log p$ bits. Our construction is valid for the regime $n = O(d/\log_p d)$, and any reconstruction threshold $k \ge 2$; in fact, the scheme attains perfect security (statistical distance exactly zero) against single-block leakage. Our technique is a partial fraction nondegeneracy argument that exploits the distinct poles of the rational iterates.

Fixed‑weight sampling is a core primitive in many post‑quantum schemes, including the HQC key encapsulation mechanism. An early implementation of fixed-weight vector sampling in HQC was shown by Guo et al. (CHES 2022) to suffer from a timing side-channel vulnerability, leading to complete key recovery. This timing side-channel was fixed in the current HQC version, however, power side-channel leakage is not addressed. In this work, we demonstrate that fixed-weight vector sampling in HQC is vulnerable to power side-channel attacks and present two practical attacks. First, we construct a power-based distinguisher targeting the support-vector generation and employ the strategy developed by Guo et al. (CHES 2022) to recover the shared key. Our attack recovers the key with a 100% success rate using 900,000 distinguisher calls. On these grounds, we evaluate hiding countermeasures based on dummy operations and find that they linearly increase the trace requirement for a successful distinguishing attack by the number of dummy operations. Second, we target a masked software implementation of the fixed-weight vector sampling in HQC and demonstrate a single‑trace attack on the support conversion that recovers the secret key, again with a success rate of 100%. We then discuss leakage attribution, specifically how shares are unintentionally recombined. Finally, we investigate how hiding techniques such as bitslicing, shuffling, and dummy operations can enhance the security of the implementation. In particular, the use of shuffling can lead to a complete prevention of our attack. Our results show that fixed-weight vector sampling of HQC is highly susceptible to power side-channel analysis. In particular, our results highlight that a combination of masking and hiding is required to effectively protect the implementations.

EME is a parallelizable encrypt-mix-encrypt wide-block construction proved secure for m-block messages only in the range m <= n, where n is the block length of the underlying block cipher. Halevi and Rogaway justified this restriction by giving a splicing distinguisher for m >= n + 2, but left open the first excluded length, m = n + 1. We resolve that boundary for the direct, unrefreshed EME-core extension beyond its specified m <= n domain: the original formulas are applied to m = n + 1 blocks while continuing to use the same global mask. Under one fixed tweak, two encryption queries and one decryption query distinguish this extension from a random permutation. The result applies to this unrefreshed extension; refreshed variants such as EME* and IEEE EME2-AES are outside the scope of the distinguisher. The boundary attack is not a shortened form of the known zero-sum attack. At m = n + 1, the non-first coefficients have no nonempty zero-sum; instead, they form a basis and therefore uniquely represent the coefficient 1. Combining this representation with the special equation for the first mixed block gives the cancellation needed for a ciphertext splice. The same mechanism applies to Sarkar's EMME framework: when multiplication by the field element x is replaced by a linear operator psi with degree-n minimal polynomial, the minimal-polynomial relation supplies the corresponding operator identity. EME* and IEEE EME2-AES avoid this setting by refreshing the mask at chunk boundaries. The attack identifies the algebraic obstruction that such refresh steps avoid.

Dilithium, as a quantum-secure digital signature standard in FIPS 204, has received widespread attention for its physical implementation security. NIST selected Dilithium's randomized signing mode as the default, which can mitigate the severe physical attacks that exploit the deterministic signing mode. However, the physical attack resilience of randomized signing mode is currently an open question. In 2024, Krahmer et al. demonstrated a key-recovery attack against randomized Dilithium by exploiting fault injection. The skipping fault correction attack in Krahmer et al.'s work benefits from operational simplicity in practical settings. Nevertheless, it uses a full-rank collection strategy, requiring several effective faults equal to the number of key coefficients. By developing an optimized skipping fault correction attack, we prove the non-necessity of the full-rank collection strategy. Theoretically, we derive the minimum number of faults M_min and a key-dependent trend for reliable key recovery. For the M_min, we obtain it by analyzing secret coefficient coverage and unique bounded solution probability under the random-row model. Lemma~1 guarantees high-probability coverage of all coefficient positions, while Lemmas~2--~4 and Theorem~1 prove sufficient uniqueness conditions. Furthermore, we instantiate M_min of Dilithium for security levels L2, L3, and L5 with a high key recovery probability. For the key-dependent trend, Remark~2 reveals that the secret key with a larger number of $\pm \eta$ in the coefficients (denoted as $\mathrm{abs_\eta}$) tends to be easier to recover, offering a feature-based insight into recovery efficiency. We formulate the key recovery of Dilithium as an MILP problem and propose an MILP model. Based on the MILP model, we design algorithms of adaptive skipping fault correction attacks for plain and shuffling settings to recover the key with fewer faults. In our experiments, we explored factors influencing key recovery success, confirming theoretical predictions that success rates increase with more faults and larger $\mathrm{abs_\eta}$. Compared to Krahmer et al.'s work, our improved attack reduces the required number of faults for recovering the private key. Specifically, for the plain setting, fault reductions were 25.9% (L2), 16.2% (L3), and 25.6% (L5). Similarly, for the shuffling setting, reductions were 25.6% (L2), 13.5% (L3), and 26.3% (L5).

The proliferation of embedded cryptographic devices in the Internet of Things (IoT) ecosystem has elevated the importance of physical security assessments. Although traditional Fault Analysis (FA) methods exhibit significant effectiveness in cryptographic key recovery, their practical application is heavily constrained by rigid mathematical requirements, the demand for precise physical fault injection, and a sensitivity to measurement noise. To address these limitations, this paper proposes Deep Learning-based Fault Analysis (DLFA), a comprehensive attack framework. By decomposing cryptanalysis into tailored feature engineering and neural network classification, DLFA successfully unifies four prominent fault models (i.e., Differential (DFA), Statistical (SFA), Statistical Ineffective (SIFA), and Persistent Fault Analysis (PFA)) under a single data-driven paradigm. Extensive physical evaluations on the SAKURA-G FPGA implementing AES-128 demonstrate that DLFA reduces the data complexity and computational time overhead compared to classical algebraic solvers. More crucially, DLFA exhibits sustained analytical stability against severe physical injection noise, relaxing the stringent hardware requirements for attackers. Finally, we employ the Integrated Gradients (IG) principle to conduct a quantitative attribution analysis, proving that the neural networks autonomously learn valid cryptographic leakages rather than overfitting to experimental artifacts.

Isogeny-based cryptography relies its security on the hardness of the supersingular isogeny problem: finding an isogeny between two supersingular curves defined over a quadratic field extension of $\mathbb{F}_{p}$. The Delfs-Galbraith algorithm is one of the most efficient procedures for solving the supersingular isogeny problem with a time complexity of $\mathcal{\tilde{O}}(p^{1/2})$ operations. The bottleneck of the Delfs-Galbraith algorithm is the so-called subfield curve search (i.e., finding an isogenous supersingular elliptic curve defined over the prime field), which determines the time complexity of the aforementioned algorithm. Given that, for efficiency, most recent isogeny-based constructions propose using finite fields with field characteristics equal to $p = 2^a \cdot f - 1$ for some positive integers $a$ and $f$. This work primarily focuses on primes of that particular form and presents two heuristic algorithms for finding subfield curves with a time complexity of $\mathcal{O}(p^{1/2})$ operations and a memory complexity polynomial in $\log_2{p}$. We show how to adapt these algorithms to primes of the form $p = d^a \cdot f - 1$ with $d$ being a small integer, with a particular focus on $d=12$. We provide concrete time-complexity bounds for both kinds of primes $p = 2^a\cdot f - 1$ and $p = {12}^a \cdot f - 1$. Our algorithms exploit the existence of large $d^a$-torsion points and extend the subfield root detection algorithm of Corte-Real Santos, Costello, and Shi (Crypto, 2022) to a projective scenario where one can work with the curve coefficients instead of their explicit j-invariants. We highlight that our algorithms easily extend to primes of the form $p =2^a \cdot f + 1$, $p =4 \cdot d_1 \cdot d_2 \cdots d_n - 1$ for small odd primes $d_i$'s, and primes such that $p - 1$ and $p + 1$ are $B$-smooth for some integer $B$.

We present a concrete bit-operation cost model for solving multivariate quadratic systems with XL using Wiedemann linear algebra, and Berlekamp-Massey sequence recovery. Following the CryptAttackTester methodology, we implement XL in a circuit-oriented model and derive closed-form cost formulas for the XL, Wiedemann, and Berlekamp-Massey steps. We instantiate the model for GF(2), GF(31), and GF(256), including baseline, constant-coefficient, and bucketed matrix-evaluation variants. Experiments on small parameter sizes show that the formulas accurately predict the circuit costs, while asymptotic analysis confirms convergence to the expected leading constant factors determined by the underlying field arithmetic. We apply the resulting estimates to Fukuoka MQ Challenge instances and to multivariate candidates from the NIST additional-signature process, providing a unified bit-operation comparison of direct Wiedemann-XL costs across several MQ-based schemes.

This paper presents a side-channel attack on the Linear Equivalence Signature Scheme (LESS) v2.0. LESS derives its security from the Linear Equivalence Problem and was evaluated as a candidate during Round 2 of the NIST post-quantum cryptography standardization process. LESS secret keys are used to generate monomial matrices, which are stored efficiently in two one-dimensional lists: the permutation list and the coefficient list. Recovering the secret monomial matrices is sufficient to forge signatures, as they are the values actually used during signing. We propose a profiled, single-trace horizontal attack on LESS key generation that recovers the full secret monomial matrices. First, the monomial coefficients that are multiplied by the dense part of the public generator matrix are recovered via power analysis of the matrix multiplication function. Next, we attack the reduced row echelon form function to recover the permutation list. We then complete the attack algebraically via two independent paths: the Primary attack and the Secondary attack. The Primary attack uses only the recovered coefficients in matrix multiplication together with their permutation positions, and a known parity-check matrix equation. The Secondary attack is an alternative that relies on another algebraic relation between the secret key and the public key and uses all the recovered values. We validated our attack on an ARM Cortex-M4 microcontroller. On the NIST Category 1 parameter set, the Primary attack achieves a 99.1% exact-recovery rate and the Secondary attack achieves a 99% exact-recovery rate, over 6000 independent keys. We also analyze potential countermeasures and show that independently shuffling the row processing order within each column reduces the success rate of our attack to negligible levels, providing protection against the specific attack vector demonstrated in this paper.

Edge-AI Internet-of-Things (IoT) devices increasingly perform local neural-network inference in physically accessible environments. While prior physical attacks on neural networks mainly focus on model extraction, parameter recovery, or intermediate computation, the final decision stage remains insufficiently studied. This paper investigates the physical security of ArgMax-based decision logic, which converts output scores into the final class label in many edge-AI deployments. We show that a sequential ArgMax implementation can expose deterministic power-trace patterns associated with loop iterations and conditional maximum updates. These patterns can be used to recover decision-related timing information and guide voltage fault injection, enabling targeted redirection of the final inference result. We implement the attack on a real MCU-based edge-AI platform using an STM32F407VG device running an X-CUBE-AI-deployed convolutional neural network. Under dual-point voltage fault injection, the unprotected ArgMax achieves targeted redirection to all non-original classes in the MNIST case study. To mitigate this threat, we propose Shuffled-ArgMax, a lightweight software-level defense that combines randomized traversal, loop-integrity checking, and redundant decision verification. Experimental results show that, under the same attack setting, Shuffled-ArgMax reduces the targeted redirection success rate from 56.92% to 0%, demonstrating strong resilience against targeted misclassification attacks. Meanwhile, it introduces only a small end-to-end runtime overhead relative to the complete DNN inference.

We study the syndrome decoding problem (SDP) in the presence of side information. The SDP asks, given a binary parity-check matrix $\mathbf{H}$ and a syndrome $\mathbf{s}$, to find a low Hamming weight binary error $\mathbf{e}$ such that $\mathbf{H} \mathbf{e} = \mathbf{s}$ over $\mathbb{F}_2$. Recent work (Cayrel et al., Eurocrypt '21) exploits a fault injection attack to reveal syndrome entries over the integers, referred to as perfect hints. Subsequent works considered side-channel scenarios to reveal similar, but noisy, information (approximate hints). Both types of hints have been shown empirically to allow for solving the SDP once enough of them are available. However, fundamental questions about the impact of these hints on the hardness of the SDP, such as thresholds for a collapse into the polynomial-time regime or how to exploit arbitrary amounts of hints, remain open. In this work, we show that both types of hints effectively allow one to transform the SDP instance into a soft-decision decoding instance. We then adapt Information Set Decoding (ISD) algorithms, the best known technique to solve generic SDP instances, to this setting. In contrast to previous work, we obtain non-trivial speedups for any amount of available hints, interpolating smoothly between the complexity of standard ISD (no hints) and polynomial time (sufficient hints). Furthermore, our practical simulations show that Hint-ISD achieves the polynomial-time regime generally under fewer hints than previous approaches. We then provide an explicit bound on the number of hints required to reach the polynomial-time regime. This bound confirms earlier practical observations that higher error weights, such as those found in the McEliece cryptosystem, exhibit higher resistance against hint exposure than schemes using smaller error weights, such as HQC.

AESpolyW, proposed at IEEE S&P 2026, is a wide-block encryption mode designed to exploit instruction-level parallelism between AES and polynomial hashing. Although AESpolyW achieves high throughput and outperforms HCTR2 and EME on most tested platforms, we show that AESpolyW does not achieve strong tweakable-PRP security. Our attack requires only two encryption queries and one decryption query, and distinguishes AESpolyW from an ideal tweakable permutation with advantage close to one. We further show that this weakness directly compromises Encode-then-Encipher authenticated encryption (AE) instantiated with AESpolyW. For the appended-zero, prepended-zero, and inserted-zero variants, we construct fresh ciphertext forgeries that are accepted with probability one. The same attacks also yield chosen-ciphertext plaintext disclosure and complete IND-CCA breaks. These results are structural and do not rely on any weakness of AES, the polynomial hash, or PHASH.

Deep Learning-based Non-profiled Side-Channel Analysis (DL-NSCA) enables automatic feature extraction without a profiling device, but existing approaches mainly target non-linear operations, requiring prior knowledge of the algorithm's unique non-linear structure and computable non-linear intermediate values. These limit applicability in analyzing proprietary or undisclosed implementations and in settings where plaintext/ciphertext are masked by unknown randomness (e.g., tweaks or nonces). We observe that linear operations are fundamental as common cryptographic primitives appearing at the beginning or end of algorithms in conjunction with the secret key, and are widely used to mask sensitive input/output. Motivated by this observation, we propose a new DL-NSCA perspective that targets the outputs of linear operations, referred to as symmetric leakage, to enable cross algorithm attacks. The main limitation of the prior distinguisher lies in their reliance on a simplistic correspondence between deep learning metrics and side channel information. This leads to two issues: the effectiveness of the distinguisher varies significantly with the chosen training epoch, and a wrong key inducing a negative correlation may be indistinguishable from the correct key under symmetric leakage. To address this, we provide a formal algebraic characterization of the relationship between the structure of the leakage function and the number of maxima given from the distinguisher. Guided by this theory, we propose a new distinguisher, VS-GBA, an epoch-invariant distinguisher that interprets SCA information from deep learning metrics and approaches the theoretical optimum. It is applicable to both the asymmetric leakage and symmetric leakage through a structure-aware screening criterion. Our experiments show that, on the high-noise 32-bit ARM Cortex-M4 device, when the core operations are protected, asymmetric leakage analysis fails to recover the keys for all three evaluated algorithms within the maximum trace budget (GE=70 for masked AES, GE=27 for masked PRESENT, and GE=66 for masked ASCON). In contrast, VS-GBA, which targets symmetric leakage, recovers the key with a 100% success rate using 8,000, 8,500, and 16,000 traces, respectively. Furthermore, we present a DL-NSCA attack on XTS-AES (NIST SP 800-38E), extending DL-NSCA to scenarios where plaintext/ciphertext is masked by a secret tweak.

NAT is widely assumed to conceal the private IP addresses of internal clients from both off-path attackers within the same LAN and external observers. We show that this assumption does not hold. We present a novel off-path deanonymization attack that infers the private IP addresses of NATed clients engaged in active TCP connections with a remote server. Our attack builds on known NAT behaviors in real-world Wi-Fi routers, such as port preservation, insufficient reverse-path validation, and the absence of TCP window tracking, previously shown to enable off-path TCP hijacking. By inferring a NATed client’s private IP address, our attack defeats the anonymity commonly attributed to NAT, creating a persistent privacy leak that enables deanonymization and user profiling. We validate the attack on 6 Wi-Fi routers and 2 real-world Wi-Fi networks using SSH and HTTP/HTTPS traffic over TCP, finding that 5 of the 6 routers and both networks are vulnerable.

This work introduces a polynomial-time attack on the signature scheme Facto-DSA. We provide an implementation that breaks all proposed parameter sets, including the largest, in under one minute on a standard laptop. These results question the suitability of multivariate polynomial factorization as a foundation for robust cryptographic schemes.

We present key-recovery attacks on the constructions of TALUS (Kao and Chang), a threshold ML-DSA (FIPS 204) construction available on arXiv and scheduled for presentation at the NIST Threshold Call Preview Talks Round 2 (TCPT-2, https://csrc.nist.gov/events/2026/tcpt). For TALUS-MPC, which is claimed EUF-CMA secure against an adversary corrupting up to $T − 1$ parties, we show the claim is false via two independent attacks, both exploiting the same root cause. TALUS-MPC uses Feldman commitments that apply the public matrix A to secret values: key shares of $s_1$ during key generation, and contributions to the nonce $y$ during signing. Since A is left-invertible in every ML-DSA parameter set, these images are invertible by Gaussian elimination, with no lattice problem to solve. A passive observer recovers all key shares of $s_1$ directly from the key-generation broadcast, and independently recovers the aggregate nonce y from the signing broadcast, which then yields $s_1 = c^{−1} \cdot (z − y)$ from a single signature. For TALUS-TEE and TALUS-MPC, we identify a persisting flaw: the rejection-sampling check that protects the error term $s_2$ in standard ML-DSA has been removed. Each signature leaks a noisy linear equation in $s_2$; applying least-squares recovery over the cyclotomic ring - an instance of LWE without modular reduction - recovers the full secret from a few hundred million signatures. The sample counts we derive are not optimized and we believe exploiting the bounded noise structure and lattice-reduction techniques would reduce them significantly, but we focus on establishing the structural flaw.

The assumed hardness of the Shortest Vector Problem in high-dimensional lattices is one of the cornerstones of post-quantum cryptography. The fastest known heuristic attacks on SVP are via so-called sieving methods. While these still take exponential time in the dimension $d$, they are significantly faster than non-heuristic approaches and their heuristic assumptions are verified by extensive experiments. $k$-Tuple sieving is an iterative method where each iteration takes as input a large number of lattice vectors of a certain norm, and produces an equal number of lattice vectors of slightly smaller norm, by taking sums and differences of $k$ of the input vectors. Iterating these ''sieving steps'' sufficiently many times produces a short lattice vector. The fastest attacks (both classical and quantum) are for $k=2$, but taking larger $k$ reduces the amount of memory required for the attack. In this paper we improve the quantum time complexity of 3-tuple sieving from $2^{0.3098 d}$ to $2^{0.2846 d}$, using a two-level amplitude amplification aided by a preprocessing step that associates the given lattice vectors with nearby ''center points'' to focus the search on the neighborhoods of these center points. Our algorithm uses $2^{0.1887d}$ classical bits and QCRAM bits, and $2^{o(d)}$ qubits. This is the fastest known quantum algorithm for SVP when total memory is limited to $2^{0.1887d}$.

In this article, we present a side-channel attack on unprotected implementations of the Berlekamp-Massey (BM) algorithm in both the reference implementation and an embedded implementation of the Classic McEliece KEM, where BM is used during decapsulation as part of the decoder implementation. We conduct a chosen cipher-text key recovery attack that exploits the power consumption of the BM, which is highly dependent on the secret Goppa support elements. We exploit the relation between plain-texts of small Hamming weight, secret elements in the Goppa support and power traces using an efficient Template Attack. Our method completely recovers the secret Goppa support for the first parameter set of the Classic McEliece KEM using a single attack trace per secret coefficient. The entire support can be recovered in less than 50 seconds. The remaining part of the secret, the Goppa polynomial, is recovered using Kirshanova and May’s algorithm (“Breaking Goppa with hints”) in less than 1 minute. Our experiments are performed on the reference implementation using the ChipWhisperer-Lite board platform with the ARM Cortex-M4 microcontroller. We also provide further insights into other implementations as well as an overview of possible extensions and limitations of our attack.

HAWK is a signature scheme that was introduced in 2022, and uses the lattice isomorphism problem (LIP) as a basis for post-quantum cryptography. In this work, we describe a classical algorithm that recovers the HAWK secret key in probabilistic polynomial time, assuming four number-theoretic heuristics. The reduction from the rank-2 module-LIP instances underlying HAWK to nrdPIP (Eurocrypt '25) is central to our algorithm. At a high level, we first conjugate the HAWK public Gram matrix $G$ by a random lower-triangular unimodular matrix $U$ with `short' entries, forming a new Gram matrix $G':=U^\ast GU$, and then test whether the $\mathcal{O}$-nrdPIP instance attached to $G'$ is unusually easy. In particular, for a non-negligible proportion of such instances $G'$, one can use the Lenstra-Silverberg algorithm to solve the corresponding $\mathcal{O}$-nrdPIP instance using a subfield approach. By resampling $U$ until such an instance is uncovered and solved, which can be seen as `re-randomising' the $\mathcal{O}$-nrdPIP instance whilst fixing the corresponding module-LIP instance, we are then able to recover a valid HAWK private key. At the time of writing, we do not claim that HAWK is broken, as we have not yet verified these heuristics experimentally. On the other hand, these heuristics seem to be very plausible, and we hope to be able to verify this in the future with an implementation of our algorithm. Update (30/06): Following discussions with Daniel Apon and Markku-Juhani Saarinen, we acknowledge that Heuristic 4 is insufficient to conclude that the main algorithm runs in polynomial time, and in fact the main algorithm appears to run in super-polynomial time. This mistake originates from the count of ideals of norm $q’$ in $\mathcal{O}_F$: one must include fractional ideals in this count, of which there are many. We note as an aside that Heuristics 1-3 have been independently experimentally verified. We would also like to thank the HAWK team and Alice Pellet-Mary for their responses to our work.

We show that the runtime complexity of the attack described in \emph{``Cryptanalysis of HAWK: a Guessing Game''} is much higher than originally claimed by its authors, and the attack is unlikely to pose a threat to HAWK's security in its present form. The attack algorithm had not been implemented before this work; the polynomial-time running-time claim was based on four ``plausible heuristics''. Our experiments and implementation data point to a super-polynomial class-number obstruction, consistent with exponential-scale growth. The experiments also helped to identify faulty ``Heuristic 4'' as the source of the observed computational wall when scaling dimension $n$. The authors of Guessing Game have acknowledged our findings. To make the argument more universal, we also offer a machine-checked conditional reduction from explicit assumptions that shows the complexity to be at least super-polynomial. In terms of methodology, our work demonstrates the role of powerful AI tools in contemporary cryptanalysis -- the sudden feasibility of rapid exploration and trial implementation of advanced attack techniques. A public research artifact contains all source code and datasets to reproduce our results.

Any signature scheme has to protect its secret key via some properly chosen, secret randomness. We show that, for the lattice signatures \textsf{HAWK}, Falcon and ML-DSA, even minimal leakage of this randomness suffices for secret key recovery. In particular, leaking either the Hamming weight or a single bit of any randomness coordinate allows an attacker to infer the sign of that coordinate. This corresponds to learning $\textrm{sign}(\langle \mathbf b, \mathbf w \rangle)$, where $\mathbf b$ is the secret key and $\mathbf w$ is public. We model key recovery from such sign information as an instance of Learning a Halfspace. This well-studied problem from learning theory provides a rich solution machinery, which we adapt for the cryptanalysis of lattice-based signatures. As a first main result, we resolve the open problem of recovering the secret key in \textsf{HAWK} from sign leakage. At the 128-bit security level and in the noise-free setting, we recover the secret key from only 30 signatures in 10 minutes. As a second main result, we recover the secret key in Falcon via sign leakage from only 100 signatures in under a minute. In comparison to existing attacks, this reduces the number of required signatures by a factor of $250$. As a third result, we show the first ML-DSA secret key recovery from sign leakage, which requires 190,000 signatures and completes within seconds. In comparison to existing ML-DSA attacks, we require a comparable amount of signatures, but utilize a less restrictive leakage model. In addition, our attack is alarmingly noise-tolerant, succeeding with up to 35\% noise for \textsf{HAWK}, 30\% for Falcon, and 35\% for ML-DSA, albeit requiring significantly more signatures in the noisy case.

Given a set of $k$-sparse linear equations over a ring $R$, we give algorithms to determine whether the right-hand sides are random or have a secret assignment planted with noise. For a parameter $k/2\leq l\leq n$, we give a spectral method to solve this problem in $\widetilde{O}\left(\binom{n}{l}\lvert{R}\rvert^l\right)$ time except with probability at most $n^{-\Omega(l)}$, provided the number of samples is roughly at least $\left(\frac{\lvert{R}\rvert n}{l}\right)^{k/2}$. This attack generalizes the Kikuchi method described by Wein et. al. (Journal of the ACM 2019) for $\mathbb{Z}_2$ to (commutative) rings of any finite size. We also give a simpler algorithm with better runtime than the spectral method and better sample complexity when $\lvert{R}\rvert =\omega(n/l)$. As a consequence, we obtain new sample-time tradeoffs for the decision problem of sparse LWE, sparse LPN over higher modulus $q$, and in general the distinguishing random vs planted $\mathbb{Z}_q$-linear equations for a large class of noise distributions. Our results imply a tightness of the hardness claims of Jain, Lin, Saha (Annual International Cryptology Conference, 2024) for sparse LWE.

In 2022, Cotan and Te{\c{s}}eleanu proposed an alternative RSA construction in which the modulus takes the form $N = pq$ and the exponents $(e,d)$ are linked through $ ed -1\equiv 0 \pmod{\psi_n(N)},\ n \geq 2, $ where $ \psi_n(N) = \frac{(p^n-1)(q^n-1)}{(p-1)(q-1)}. $ Their scheme was subsequently examined by Nitaj et al. at Africacrypt 2024, who demonstrated that the system becomes vulnerable when the secret exponent $d$ is less than $N^{0.292(n-1)}$. In the present work, we extend this line of cryptanalysis by developing a lattice-based technique targeting the same key congruence. Our approach succeeds in recovering the private parameters even for secret exponents greater than $N^{0.292(n-1)}$, thereby surpassing the previously known boundary. The factorization of $N$ is achieved in polynomial time, provided that its divisors $p$ and $q$ share a suitably portion of their least significant bits.

Lattice-based cryptography is one of the most promising candidates for post-quantum cryptography. Decryption Failure Rate (DFR) is a critical metric for the correctness and security of lattice-based public-key encryption (PKE) schemes. At present, most DFR evaluation methods for lattice-based PKE with message encoding rely on oversimplified assumptions, rough approximations, and fail to fully exploit the geometric structure of special encoding lattices, resulting in loose or inaccurate bounds. To address these limitations, this paper proposes a refined DFR evaluation framework for lattice-based PKE with message encoding, focusing on two mainstream decoding paradigms: Maximum Likelihood Decoding (MLD) and Bounded Distance Decoding (BDD). For MLD-based schemes, we precisely characterize the minimal vectors of Barnes-Wall (BW) lattices and derive tighter union bounds by leveraging the encoding lattice structure. For BDD-based schemes, we introduce a noncentral chi-squared distribution based method to model mixed noise composed of Gaussian and discrete components, avoiding the inaccuracy of conventional pure Gaussian approximations. Furthermore, we extend the framework to algebraic-lattice-based PKE by analyzing the variance and correlation of polynomial product coefficients, and propose a weighted chi-squared distribution with saddlepoint approximation to handle correlated coefficients. We validate the proposed methods on representative lattice-based PKEs including CNTR, \scloud, and CNTR-Prime, resulting in the DFR upper bound approximate 15 bits decreasing for CNTR, 1 bit decreasing for \scloud, and at least 84 bits increasing for CNTR-Prime. The experimental results show that our refined analysis yields significantly tighter and more accurate DFR bounds compared with traditional approaches, providing a reliable theoretical basis for parameter selection and security certification of practical lattice-based cryptosystems.

The Large Norm attacks of Ducas-Espitau-Postlethwaite (CRYPTO 2023) on the $\mathsf{ISIS}$ problem have shown that small values of $q$ can be used to recover short solutions. This was applied to Falcon and Mitaka. Two issues are left, however. First, the cost model of the attack oversimplifies the BDGL sieve; it does not account for how long vectors will be distributed and treats two dependent probabilistic events as if they were independent, which results in an overestimation of the attack cost. Secondly, the analysis only deals with the $\ell_2$ norm and has not yet been extended to $\mathsf{ISIS}^\infty$, which underlies Dilithium-type systems. This work addresses both issues. First, we extend the cost model of the Large Norm attack by including the principal sieve length distribution into our estimation of success probabilities. In addition, we use a joint probability rather than an approximate factor where possible. Since all other parts of the original $\theta$ convolution framework are reused, the extension is relatively minor. We reduce the attack cost of Large Norm on Falcon-256 by a $\approx\!11\times$ cheaper model, and successfully forge a Mitaka-512 signature in $\approx\!4.5$ seconds at a higher success rate. We additionally introduce a closed-form $\ell_\infty$ variant as a Z-shape attack against Dilithium-type $\mathsf{ISIS}^{\infty}$ at small-to-moderate modulus, which succeeds in $\le\!1.6$ seconds across three presets.

With the standardization of post-quantum digital signature schemes and their increasing deployment in security critical applications such as firmware authentication and software distribution, implementations are expected to operate in physically accessible and potentially hostile environments. Consequently, considerable effort has been devoted to protecting these schemes against a variety of attacks, including timing sidechannel attacks. However, evaluating their resilience against fault injection attacks remains equally important. Previous differential fault analysis (DFA) attacks on lattice-based signatures have primarily targeted intermediate values during signing and often relied on assumptions regarding rejection sampling or multiple fault injections. In this work, we demonstrate that the challenge sampling procedure itself constitutes a practical attack surface. Specifically, We present fault attacks against the challenge sampling procedures of deterministic ML-DSA, a NIST-standardized signature scheme, and HAETAE, a KpqC-selected signature scheme, showing that a single faulted signature is sufficient to recover the secret key required for signature forgery. To the best of our knowledge, this is the first fault attack on HAETAE achieving secret-key recovery that enables the generation of valid forged signatures. Our attack model of ML-DSA does not require direct access to faulted challenges. Using only public information, we identify intended fault injections and distinguish them from unintended fault outcomes. We evaluate the method through simulation and practical fault injection, achieving a 100% identification rate for intended faults. We further propose a countermeasure for the identified vulnerability

The Estonian i-voting experience, since 2005, is probably the richest to analyze; yet it still holds a collection of noteworthy threats as stated in the 2025 Estonian Academy of Science (Cyber Security Committee) report. This paper first gathers vulnerabilities discovered and improvements/fixes introduced to the Estonian i-voting system, IVXV, in the years 2023-2025; then presents a thorough analysis of its status, as of October 2025 version, with exploits pointed out and solutions suggested. In our review, we shade some light on automated formal verification research attempts that went unnoticed by the Estonian research community; we also notice that, since 2024, forcing a time gap between multiple votes could prevent Perreira revoting attack for verifying voters. Although the problem of not authenticating the Voting Application (VA) is not new and was discussed in the Estonian report, we warn of possible threats that have been missed. We show how new technologies like anonymous online auctions between candidates and hardware execution attests for conditional payment can be used to orchestrate an automated large-scale attack where an encrypted vote is generated online without the vote sellers even knowing who they’re voting for; adding fresh randomness to each ballot enables the attack to pass duplicate ballots anomaly check if existed. We agree that compromised voters’ credentials is a known threat to all i-voting systems; however, it only gets severe in the Estonian case where any application can cast a vote. Finally, we discuss the approaching quantum threat and recommend protecting election data from the possible HNDL (Harvest Now Decrypt Later) privacy threat by encrypting ballots, while residing on the system servers, using a quantum secure symmetric key algorithm (AES). The appendices contain also two suggested solutions for insiders' threat.

We show how to improve rank-metric solvers when certain side information (hints) about the secret is available. Concretely, we adapt the kernel search algorithm for MinRank and the GRS algorithm for the Rank Syndrome Decoding problem when some entries in the rank decomposition of the error matrix are known. This setting is motivated by side-channel leakage and cryptographic applications: Mirath and RYDE, two signature candidates in the NIST post-quantum competition, rely on these problems and employ secret keys in this decomposed form. As a main technical ingredient, we give an optimal procedure for guessing a subspace containing the row space of a systematic matrix given only partial knowledge of its entries. Further, we describe a profiling side-channel attack on the reference implementation of Mirath to demonstrate the plausibility of obtaining such hints.

Multivariate quadratic (MQ) signature schemes such as MAYO are among the leading candidates for post-quantum digital signatures, with security relying on the hardness of solving systems of multivariate quadratic equations. In this paper, we present a fault-assisted man-in-the-middle attack targeting the key-generation procedure of MAYO. Specifically, we target the computation of the public key, which is represented as a system of $m$ quadratic polynomials. Following the MAYO specification, each polynomial is associated with a matrix representation composed of the components $P_i^{(1)}, P_i^{(2)}$, and $P_i^{(3)}$. Our attack injects a single fault during the computation of the $P_i^{(3)}$ component of the public key, causing the resulting public key to satisfy a simplified linear relation involving the secret oil matrix. From the faulty public key, we derive an overdetermined linear system over $F_{16}$ that enables complete recovery of the secret oil matrix. Unlike previous fault attacks that primarily target the signing algorithm, our attack targets the key generation procedure to recover the oil secret and exploit the recovered oil secret to reconstruct the legitimate public key corresponding to the victim's secret key i.e correcting the public key. This allows the adversary to transparently participate in the communication as a man-in-the-middle, producing valid signatures on behalf of the victim. We evaluate the attack on a fault simulated MAYO implementation and demonstrate successful recovery of the oil matrix from a single faulty key-generation execution followed by public key correction.

We give a new algorithm to attack RSA with small private exponent, when some partial information of $p + q$ is given. The algorithm is a very simple modification of original Wiener’s attack with continued fractions, and allows us to factor $n$ whenever $d<n^{(1+\delta)/2}$ if we know a $δ$-fraction of the most significant bits of $n$. The algorithm is unconditional, which is not the case in previous improvements that use Coppersmith method. As a simple example, our algorithm can be applied to break any cryptosystem with modulus $n$ of $512$ bits and $d < n^{0.3}$, given an improvement in the original. attack of Wiener.

Every formally verified system embeds a verification boundary: the interface between code with machine-checked proofs and code that is trusted without them. We study what happens when this boundary is not communicated clearly. Through a case study of CE Labs's libcrux and hpke-rs cryptographic libraries, we present thirteen vulnerabilities that escaped formal verification. Nine reside in unverified code, including a cross-backend endianness bug that caused real decryption failures in Signal's post-quantum ratchet, a missing mandatory X25519 validation, nonce reuse via integer overflow, and two FIPS~204 specification violations in the ML-DSA verifier. Four reside in formally verified specification and proof code: in ML-KEM, a wrong decompression constant, a missing inverse NTT, and a false serialization proof; in ML-DSA, a wrong multiplication specification that renders axiomatized AVX2 proofs unsound. From these findings, we develop a taxonomy of five verification boundary failure types, a lightweight auditing methodology for detecting them, and a comparative analysis with AWS's verified libcrypto. The same failure types arise in both projects, but their management---through systematic documentation, proof execution in CI, and clear scope communication---varies significantly. We call the gap between verification claims and verification reality "verification theatre", and propose concrete practices for closing it.

Algebraic attacks are an important class of cryptanalytic techniques for Poseidon/Poseidon2 and Neptune. These designs use partial S-box activation in internal rounds to reduce arithmetization cost, and this structure makes linear subspace trails an important technique in their algebraic analysis. If the state at some internal round is restricted to a suitable linear subspace, then several subsequent internal rounds can be linearized and therefore do not increase the degree of the polynomial system used to model the attack. Recent studies have explored such linearization and round-skipping techniques for algebraic attacks on Poseidon/Poseidon2 and Neptune. These techniques reduce the complexity of attacks by lowering the degree of the polynomial systems that model the attacked rounds. Consequently, a precise bound on the length of linear subspace trails is needed for a more accurate assessment of the algebraic security of these designs. We revisit infinite and finite linear subspace trails in Poseidon-like designs. First, motivated by previous work that relates infinitely long linear subspace trails to invariant subspaces of the linear layer, we revisit this phenomenon for the Cauchy MDS matrices used in Poseidon. We further give a quantitative heuristic estimate for the probability that such invariant subspace conditions occur under random parameter choices. Second, we analyze finite linear subspace trails for partial rounds in a general setting, where the state width is $t$ and each round activates $s$ S-box coordinates. Under the rank growth condition stated in this paper, when no such invariant subspace exists, a finite trail has length at most $\lceil t/s \rceil - 1$. For Poseidon, $s = 1$, this gives at most $t - 1$ consecutive linearized internal partial rounds. For the internal linear layers of Poseidon2 and Neptune, a similar conclusion applies. More precisely, if repeated diagonal entries in the lower-right block do not give rise to infinitely long linear subspace trails, then the maximum length of a finite linear subspace trail is exactly $t - 1$. Third, for preimage attacks in sponge mode with rate $r$, capacity $c$, and digest size $d$, the available extra constraint budget is $Ec = r - \min(c,d)$. For Poseidon/Poseidon2 and Neptune, the finite trail bound and this constraint budget together determine how many internal partial rounds can be linearized in the corresponding attack model.

We show that the certificateless signature scheme [IEEE ITJ, 26852-26865, 2024] is insecure against public key replacement attack. An adversary can forge signatures for any message by replacing the signer's public key. We find the two components $\delta_A$ and $T_A$ of signature $\sigma_A=(m_A, ID_A, \delta_A, T_A)$ are not tightly bound to the target message $m_A$ and the singer's identity $ID_A$. The inherent flaw results in that the adversary can find an efficient signing algorithm functionally equivalent to the valid signing algorithm. The findings could be helpful for researchers unfamiliar with the designing techniques for certificateless signatures.

This paper is devoted to the study of the Ring-LWE-based digital signature scheme proposed by Sharafi and Daghigh, especially, the cryptanalysis of this scheme. The Sharafi and Daghigh’s scheme is inspired by the Lindner–Peikert encryption paradigm and adopts a hash-and-sign approach via the Fiat–Shamir transformation. The security claims rely on the assumed hardness of the Ring LWE and Ring-SIS problems, whose definitions and properties are well studied in last two decades. We demonstrate that this scheme is not secure and generalise our analysis to the analogous scheme in the plain-LWE setting.

HAWK is a signature scheme based on the module lattice isomorphism problem, and the only lattice-based candidate in the third round of NIST's call for additional post-quantum signatures. Its specification claims that HAWK achieves the BUFF (Beyond UnForgeability Features) security properties "as is", without applying the generic BUFF transform, citing the analysis of Aulbach, Düzlü, Meyer, Struck, and Weishäupl (PQCrypto'24). We refute this claim for HAWK exactly as specified. Several of the BUFF games let the adversary register a public key of its own choosing, yet the HAWK reference verifier performs almost no validity check on a key beyond decoding it. We exhibit degenerate "weak" public keys under which the all-zero signature verifies for all random-oracle challenges except the negligible symmetry-breaking corner case, and use them to break the three BUFF properties whose games let the adversary supply both the verification key and the signature: message-bound signatures, malicious strong universal exclusive ownership, and weak non-resignability. We trace these breaks to gaps in the BUFF proofs for HAWK of Aulbach et al.: one missing case analysis, and three steps that silently assume properties of the adversarially chosen keys as if they come from honest key generation. Honest key generation already enforces bounds that would reject our weak-key family on the attacks we exhibit. We prove that enforcing this norm floor at verification yields message-bound security for constant keys, but we make no claim that it fully restores BUFF security.

The Rank Decoding (RD) problem lies at the core of rank-based cryptography. To enable efficient constructions, several variants have been introduced, notably the Non-Homogeneous RD (NHRD) problem and the Blockwise RD (BRD) problem. The \emph{quantum} security of these systems is currently considered to be determined by the complexity of combinatorial attacks such as AGHT, PRR, and Ourivski--Johansson (OJ) attacks. However, for the OJ attack, the modeling, soundness, and relative complexities remain poorly understood, particularly for the NHRD and BRD variants, thereby limiting confidence in security claims and hindering the design of compact schemes. In this work, we refine the modelings for the OJ attack (PIT, 2002) and the Improved OJ (IOJ, IEEE TIT 2025) attack, and obtain general and tight complexities on the RD, NHRD, and BRD problems. We show that the IOJ attack rests on optimistic assumptions that do not hold in practical random decoding scenarios, and thus its advantage over OJ should be disregarded in security assessments. For the RD problem, the OJ attack remains a strong candidate for the most powerful combinatorial attack in certain parameter regions, particularly when the code dimension $k$ is small and the extension degree $m$ is large. For the NHRD problem, we show that the OJ attack is the most powerful combinatorial attack for the parameters of NH-Multi-UR-AG, yielding up to a 100-bit improvement over the adapted AGHT attack (IEEE TIT 2024), while still preserving the claimed security level. For the BRD problem, we derive complexity formulas for general block structures, resolving questions posed in prior works (Asiacrypt 2023, IEEE TIT 2025, PQC 2024). Our analysis also reveals that the OJ attack is previously underestimated by about $\gamma^2$ bits, where $\gamma$ denotes the minimum block weight. We further show that the OJ attack outperforms AGHT and PRR attacks in certain parameter regions, achieving up to a 136-bit advantage over PRR (IEEE TIT 2025). Our work advances the understanding of decoding problems in the rank metrics and reinforces the security of related cryptosystems.

Static Quantum Bit Error Rate (QBER) thresholding is the standard defense mechanism in deployed Quantum Key Distribution (QKD) systems. In noisy free-space optical (FSO) channels, however, natural atmospheric variations can camouflage short, low-intensity eavesdropping bursts, rendering fixed thresholds ineffective. This paper investigates physics-aware temporal feature engineering for machine learning-based anomaly detection in entanglement-based BBM92 QKD telemetry. A 24-dimensional feature space is computed over a 30-second sliding window, characterizing the temporal evolution and cross-observable correlations of QBER, the Bell S parameter, and photon coincidence rates. Evaluated on a simulated FSO telemetry dataset spanning 24 hours across five random seeds, static QBER thresholding achieves only 17.3% recall against blended sub-threshold attacks, while an XGBoost classifier trained on the proposed feature set achieves 96.9% recall and 97.6% precision within the same simulation framework. SHAP analysis suggests that detection is driven less by absolute error magnitude than by anomalous temporal decoupling between independent quantum observables. These results indicate that physics-aware temporal representations may improve resilience to stealthy attack strategies in simulated BBM92 environments, although validation on operational hardware and real telemetry remains an important direction for future work.

In the Linear Code Equivalence (LCE) problem, one seeks the isometry of the Hamming space $\mathbb{F}_q^n$ relating two given linear codes. This problem is considered computationally hard for any alphabet size $q \ge 5$. The matching codewords framework currently serves as the benchmark for evaluating the security of cryptographic schemes whose security relies on the hardness of LCE, such as the LESS signature scheme. The framework operates by sampling multiple low-weight codewords in both target codes until matching codewords are discovered, thereby revealing the underlying isometry. Recent advancements improved this framework by introducing a new matching algorithm to decide whether two pairs of codewords match. While this technique offers better scalability than previous approaches and has led to enhanced attacks on LESS, it remains computationally intensive for certain parameter ranges. In this work, we propose a novel method to determine if a single pair of codewords matches. Our approach is based on the Schur product by an inverse vector, $\mathbf{v}^{-1} \star \mathcal{C}$, defined component-wisely. We demonstrate that if the codes $\mathcal{C}_1$ and $\mathcal{C}_2$ are linearly equivalent and the codewords $\mathbf{v}_1$ and $\mathbf{v}_2$ match, then the resulting products $\mathbf{v}_1^{-1} \star \mathcal{C}_1$ and $\mathbf{v}_2^{-1} \star \mathcal{C}_2$ are permutation equivalent. Since the permutation equivalence problem is efficiently solvable, this provides a highly effective match-testing algorithm. By leveraging this idea, we propose several algorithms that improve the asymptotic exponent of LCE solvers for all parameters. Notably, our method reaches the optimal asymptotic exponent of the framework for a wide range of parameters. When applied to LESS, our technique substantially reduces the best-known bit complexity; for instance, reducing the security of LESS-1 from 127 to 120 bits.

We study the Lattice Isomorphism Problem (LIP) for both indefinite and definite quadratic forms, with applications to the DEFI signature scheme. By combining arithmetic and algorithmic techniques, we obtain efficient attacks on DEFIv2, a digital signature scheme based on isotropic quadratic forms. Our approach to the Decision \ Distinguishing-LIP draws on the arithmetic theory of quadratic forms, with particular emphasis on indefinite forms of dimension at least 3. We show that such forms arise naturally in the analysis of DEFI and prove that, under suitable assumptions, the genus, spinor genus, and equivalence class coincide. This structural collapse leads to a classical polynomial-time algorithm for the Decision \ Distinguishing-LIP instances obtained from DEFI. In addition, we present an efficient algorithm for recovering the secret key of DEFIv2 and demonstrate practical signature forgeries within minutes using the authors' public challenge instances. Finally, we evaluate the impact of our methods on HAWK and show that, in contrast to DEFI, they do not compromise its security.

Rocket.Chat is a group chat platform widely deployed in industries and national organizations, with over 15 million users across 150 countries. One of its main features is an end-to-end encryption (E2EE) protocol; however, no cryptographic security analysis has been conducted. We conduct an in-depth cryptographic analysis of Rocket.Chat's E2EE protocol and identify multiple significant flaws that allow a malicious server or even an outsider to break the confidentiality and integrity of the group chat. Specifically, we formally model and analyze the protocol using ProVerif under the Dolev-Yao model, uncovering multiple theoretical weaknesses and verifying that some of them lead to practical attacks. Furthermore, through meticulous manual analysis, we identify additional vulnerabilities, including implementation flaws and cryptographic weaknesses such as CBC malleability, and demonstrate how they are exploitable in practical attack scenarios. To validate our findings, we develop Proof-of-Concept implementations, highlighting the real-world feasibility of these attacks. We also propose mitigation techniques and discuss the implications of our attacks.

This paper presents the first practical end-to-end fault injection attacks on the post-quantum signature scheme PERK, based on the MPC-in-the-Head paradigm and relies on GGM tree expansions for efficient randomness generation. While GGM trees reduce memory requirements, they introduce implementation-level deviations from the theoretical model. We show that these implementation choices fundamentally alter the fault surface and enable new attack vectors that are not captured by the original security assumptions. We propose two attacks targeting distinct stages of the GGM tree construction. The first attack fixes the root seed, resulting in deterministic tree generation and enabling full reconstruction of all leaf labels. The second attack induces reuse of GGM roots via an instruction skip, allowing recovery of hidden leaf values across consecutive rounds. Both attacks require only a single fault to recover the secret key. We demonstrate both attacks end-to-end on the pqm4 reference implementation compiled with optimization levels -O3 as well as -Os for ARM Cortex-M4 microcontrollers. Our evaluation is performed on two hardware platforms (ChipWhisperer-Lite with STM32F303 and Nucleo-L4R5ZI-P) using clock glitching and electromagnetic fault injection (EMFI). We achieve success probabilities of 100% and 85% for the two attacks, respectively. Finally, we propose countermeasures for both attacks.

Canteaut et al. demonstrated that bent Boolean functions are balanced over one of two complementary affine spaces. In this work, we provide an alternative proof of this property and leverage this result to establish the existence of $1$-plateaued functions whose restrictions to these affine spaces remain bent. By incorporating these results into the secondary Maiorana-McFarland construction, we obtain bent functions that are inherently balanced over a specific affine subspace (the underlying vector subspace). Furthermore, we analyse the balancedness of these new functions under linear perturbations. Finally, two algorithms are developed to simplify the research findings. It is worth noting that the sets of vectors with even and odd Hamming weights constitute a particular case of such complementary affine spaces.

Ring Signatures allow a user to sign on behalf of an ad-hoc set of public keys, while hiding their identity inside that set. Linkable Ring Signatures (LRS) add the functionality of detecting signatures originating from the same signer. They have found many applications in anonymous transactions and e-voting. The LLRing family of linkable ring signature schemes by Hui and Chau (ESORICS 2024) is one of the more efficient LRS schemes. However, we show that it has an unlinkability vulnerability, meaning an adversary can create more unlinkable signatures than the number of secret keys they own. The vulnerability is caused by the introduction of unwanted structure to base elements used in proofs. We also find a similar attack against the Threshold Ring Referral (TRR) scheme of Ta, Hui, and Chau (Security and Privacy 2025), rendering it unsound. We show how to achieve strong linkability with logarithmic verification complexity in the pairing based setting by first reverting the unsafe construction of base elements, and by also adjusting the arguments of knowledge used in order to maintain efficiency. Concretely, by modifying the Dory argument to fit our scheme we are able to match the performance of LLRing-P. We separate the design and analysis of the scheme from the instantiation of the knowledge arguments, which helps prevent unwanted interactions between the two, and can provide easier upgrades to more efficient proof systems.

The security of most prominent code-based key encapsulation mechanisms (KEMs) relies on the hardness of the syndrome decoding problem. It is well-known that in the presence of $n$ syndromes, one gets a speed-up of roughly $\sqrt n$ for decoding a single syndrome by a technique called Decoding One Out of Many (DOOM), due to Sendrier. Modern code-based schemes like HQC and BIKE work over a polynomial ring $\mathbb{F}_2[X]/(X^n-1)$ that naturally leads to $n$ syndromes. As a consequence, DOOM-type speed-ups of $\sqrt n$ have been taking into account for the HQC and BIKE parameter selection in the single-instance setting. However, we analyse a naturally appearing multi-instance setting, where the same public key is used to derive $M$ session keys $K^{(1)}, \ldots, K^{(M)}$. Our attack goal is to reconstruct a single session key $K^{(i)}$. We show that in a BIKE multi-instance setting an attacker can construct a DOOM instance with $nM$ syndromes. In an HQC and Classic McEliece multi-instance setting, an attacker obtains $M$ syndromes. Our results show that multi-instance security of code-based KEMs degrades as a function of $M$. For KEMs designed for NIST security level 1 we drop below the desired $143$ bits for a number of session keys $M \geq 2^{34}$ ($\texttt{HQC-1}$), $M \geq 2^{11}$ ($\texttt{BIKE-1}$), respectively $M \geq 2^{21}$ ($\texttt{mcecliece3488-64}$). As a conclusion, the public keys of all three code-based KEMs should be updated regularly.

The construction of building beyond-birthday-bound secure pseudorandom functions (PRFs) from the Xor-sum of 2 pseudorandom permutations (PRPs) has been known since EUROCRYPT 1998. However, the first concrete instance was only published recently at FSE 2022: the low-latency PRF Orthros. Subsequently, at ASIACRYPT 2024, Flórez-Gutiérrez et al. proposed the general framework of ZIP ciphers, where a block cipher $E_{1} \circ E_{0}$ is used to construct the PRF $E_{0} \oplus E_{1}^{-1}$. They propose the PRF ZIP-AES, as the Xor-sum of 5 AES encryption rounds and 5 decryption rounds. They discuss differential, linear, and integral distinguishers for this construction, but provide no concrete key recovery attacks. Furthermore, they propose ZIP-GIFT as a 64-bit PRF but leave cryptanalysis as future work. In this work, we provide the first third-party analysis of ZIP-AES and ZIP-GIFT. We investigate the challenges of applying classical cryptanalytic techniques to ZIP ciphers and discuss several adaptations needed for key recovery attacks in this setting. We show differential, linear, and integral key recovery attacks for both PRFs. In particular, we adapt integral key recovery techniques for ZIP ciphers, and show how to extend differential characteristics by some rounds for key recovery using truncated differential distinguishers. Our analysis deepens the understanding of the security of ZIP ciphers but does not threaten the security claims.

HAWK is a lattice-based signature scheme candidate to the fourth call of the NIST's Post-Quantum standardization campaign. Considered as a cousin of Falcon (one of the future NIST post-quantum standards) one can wonder whether HAWK shares the same drawbacks as Falcon in terms of side-channel attacks. Indeed, Falcon signature algorithm and particularly its Gaussian sampler, has shown to be highly vulnerable to power-analysis attacks. Besides, efficiently protecting Falcon's signature algorithm against these attacks seems a very challenging task. This work presents the first power analysis leakage review on HAWK signature scheme: it extensively assesses the vulnerabilities of a central and sensitive brick of the scheme, the discrete Gaussian sampler. Knowing the output x of the sampler for a given signature leads to linear information about the private key of the scheme. This paper includes several demonstrations of simple power analysis attacks targeting this sample x with various attacker strengths, all of them performed on the reference implementation on a ChipWhisperer Lite with STM32F3 target (ARM Cortex M4). We report being able to perform key recoveries with very low (to no) offline resources. As this reference implementation of HAWK is not claimed to be protected against side-channel attacks, the existence of such attacks is not surprising, but they still concretely warn about the use of this unprotected signature on physical devices. To finish, we propose an ad-hoc technique to lower down the leakage on the identified vulnerability points. These modifications prevent our attacks on our platform and come with essentially no cost in terms of performance. It could be seen as a temporary solution and encourages more analysis on proven side-channel protection of HAWK like masking.

Session is a widely deployed decentralized messenger application that emphasizes user anonymity and privacy through end-to-end encryption. Session currently employs its own uniquely designed messaging protocol, Session Protocol V1, having migrated from the extensively studied Signal Protocol. In this paper, we conduct a comprehensive, implementation-driven security analysis of the Session Protocol V1, focusing on its 1-to-1 and closed-group communication mechanisms. Our analysis reveals two fundamental design vulnerabilities: the absence of mutual public key authentication and the lack of cryptographic bindings to monotonic sequence counters. Exploiting these weaknesses within the context of actual application environments, we demonstrate three practical attacks: an impersonation attack, a message timestamp forgery attack, and message dropping and replay attacks. These attacks allow malicious server nodes or unprivileged malicious insiders to substitute public keys, silently suppress or duplicate messages, and manipulate the perceived chronological order of conversations. The findings highlight that these exploits severely undermine the fundamental security guarantees of the messenger. Finally, we propose immediate, actionable mitigation strategies to address the identified flaws and secure the protocol against these threats.

Differential cryptanalysis is a powerful technique for attacking block ciphers, wherein the Markov cipher assumption and stochastic hypothesis are commonly employed to simplify the search and probability estimation of differential trails. However, these assumptions often neglect inherent algebraic constraints, potentially resulting in invalid trails and inaccurate probability estimates. Some studies identified violations of these assumptions and explored how they impose constraints on key material, but they have not yet fully captured all relevant ones. This study proposes Trail-Estimator, an automated verifier for differential trails on block ciphers, consisting of two parts: a constraint detector Cons-Collector and a solving tool Cons-Solver. We first establish the fundamental principles that will allow us to systematically identify all constraint subsets within a differential trail, upon which Cons-Collector is built. Then, Cons-Solver utilizes specialized preprocessing techniques to efficiently solve the detected constraint subsets, thereby determining the key space and providing a comprehensive probability distribution of differential trails. To validate its effectiveness, Trail-Estimator is applied to verify 17 differential trails for the SKINNY, LBLOCK, TWINE, and AES block ciphers. Experimental results show that Trail-Estimator consistently identifies previously undetected constraints for SKINNY and AES, and discovers constraints for the first time for LBLOCK and TWINE. Notably, it is the first tool to discover long nonlinear constraints extending beyond five rounds in these ciphers. Furthermore, Trail-Estimator’s accuracy is validated by experiments showing its predictions closely match the real probability distribution of short-round differential trails.

We introduce WayFinder, a framework for generalizing the Delfs-Galbraith and SuperSolver algorithms for the supersingular isogeny problem. Our framework extends the search for elliptic curves with an orientation by an order containing $\mathbb{Z}[\ell \sqrt{-p}]$ to more general orders, and we derive a cost model for such generalisations. Our cost model not only works in a more general context, but also provides more accurate predictions when applied to SuperSolver. We instantiate WayFinder for orders containing $\mathbb{Z}[\ell_1\sqrt{-\ell_2p}]$ where $\ell_i$ are $1$ or primes such that the modular curve $X_0(\ell_2)$ has genus $0$. We then introduce a low-storage algorithm for computing an isogeny between two oriented supersingular elliptic curves, even when the curves are oriented by distinct orders. Together, these provide an algorithm that improves on the state of the art for solving the isogeny problem, and a cost model with potential applications to parameter selection in isogeny-based cryptography.

This paper introduces a geometric framework for Q2 quantum distinguishers by combining the geometric approach to classical symmetric-key cryptanalysis with the generalized correlation extraction algorithm. Our main technical tool shows that one superposition query, followed by appropriate (unitary) change-of-basis operations, prepares a ``correlation state'' whose amplitudes are the entries of the geometric correlation matrix in the chosen basis. This yields a unified preparation-measurement template that recovers several known quantum distinguishers: (1) hidden structure detection via support constraints in Fourier-type bases (e.g., Simon, Bernstein-Vazirani, Deutsch-Jozsa), and (2) event probability deviation tests via amplitude estimation (covering standard quantizations of linear and differential distinguishers). We analyze when relevant distinguishing mass is diluted across many basis states, identify it as a cause of poor query efficiency in several recent distinguishers, and provide basis-specific mechanisms to concentrate the signal (phase-oracle row restriction in the Fourier setting; chosen-plaintext subset-state restriction in the quasidifferential setting) to restore quadratic advantage. We illustrate the framework on Fourier and quasidifferential instantiations and discuss obstacles for non-unitary integral bases.

End-to-end deep-learning side-channel attacks on public-key implementations have recently become possible even for million-sample traces. However, existing methods require large computational resources and extract only partial key shares, which means that dedicated post-processing is required to turn detected leakage into demonstrations of successful key recovery attacks. We present an end-to-end sequence-to-sequence prediction approach to recover complete 256-bit key shares from single raw traces on the SCAAML ECC datasets recently studied by Bursztein et al (TCHES 2024). Our solution combines aggressive trace compression for dimensionality reduction with a 1-D U-Net trained using Connectionist Temporal Classification loss. The key idea is to decouple detecting leakage from mapping each leakage site to the correct part of the secret: the network outputs an annotated map of the trace marking likely leakage sites, and a greedy decoder reconstructs the ordered key bits from that map. Using synthetic tasks, we show that this division of labor circumvents a fundamental problem that causes neural network architectures and training methods commonly used in side-channel analysis to struggle with massive multi-target or misaligned extraction tasks. As a result, we are able to train a single extractor that achieves high accuracy on all four SCAAML ECC datasets in a single training run that takes minutes on a single GPU. The resulting extractors are robust, essentially maintaining their performance under large misalignment (we empirically tested rotations up to \(61\%\) of trace length), while degrading gracefully under a variety of trace corruptions, and even time reversal. They transfer across key shares and datasets with little degradation and no retraining. The U-Net outputs also yield prediction maps that localize leakage along the trace prior to decoding.

The unbalanced oil and vinegar signature scheme (UOV) was proposed by Kipnis et al. in 1999 as a multivariate-based scheme. UOV is regarded as one of the most promising candidates for post-quantum cryptography owing to its short signatures and fast performance. Recently, Ran proposed a new key recovery attack on UOV over a field of even characteristic, reducing the security of its proposed parameters. Furthermore, Jin et al. generalized Ran’s attack to schemes over a field of arbitrary characteristic by exploiting the structure of the symmetric algebra. In this work, we propose a new framework for recovering the secret subspace of UOV over a finite field $\mathbb{F}_{p^e}$ by generalizing these preceding results. First, we show that a key recovery against UOV can be successfully performed using the XL algorithm by exploiting the structure of the $p$-truncated polynomial ring $R^{(p)}=\mathbb{F}_{p^e}[x_1,\dots,x_n]/ \langle x_1^p,\dots,x_n^p\rangle$. This result simplifies the description of the attacks proposed by Jin et al. by formulating them in terms of the polynomial ring, independent of the structure of the symmetric algebra. Second, we generalize this result to the polynomial rings of more general forms, namely, the $p^\ell$-truncated polynomial rings $R^{(p^\ell)}$ for any $1 \le \ell \le e$. This result is due to our description in terms of the polynomial ring and can relax the constraints on the solving degree of the XL algorithm using $R^{(p^\ell)}$ by taking a larger $\ell$. Finally, we consider performing the reconciliation and intersection attacks using the $p^\ell$-truncated polynomial rings against UOV. In particular, we consider the intersection attack using this framework, which has not been addressed in previous analyses. Based on our complexity estimation, we confirm that the optimal complexity of the reconciliation attack using the proposed framework is consistent with that of the symmetric-algebra attack by Jin et al. We further show that the intersection attack using the proposed framework outperforms the reconciliation attack against the proposed parameters of UOV and reduces the security of multiple parameters compared to their claimed security levels.

The Feistel (Luby-Rackoff) structure underlies numerous block-cipher and mode-of-operation designs, whose security is traditionally assessed via indistinguishability. For low-round Feistel constructions, a variety of classical and quantum distinguishing attacks are known. In this work, we show that such distinguishing attacks can be systematically upgraded to full plaintext recovery with essentially the same query complexity. We establish classical recovery attacks on the $2$-round Feistel under CPA and the $3$-round Feistel under CCA using only three queries, and introduce quantum-assisted forward/backward extension techniques based on Simon’s algorithm that yield recovery attacks on the $3$-round Feistel under qCPA and the $4$-round Feistel under qCCA. We further prove that the attacks extend to the Unified Feistel-Lai-Massey (UFLM) framework and therefore apply to a broad class of two-branch constructions. As a consequence, we obtain plaintext-recovery attacks on $4/5/6$-round Feistel-FK and on several practical enciphering schemes, including AEZ-core, FMix, OleF, double-decker, and docked-double-decker. Overall, our results reveal a fundamental connection between distinguishing and full plaintext recovery in low-round two-branch Feistel-type designs, in both classical and quantum settings.

The recent study has demonstrated that neural networks can serve as a navigator for an automatic search model for integral cryptanalysis with a reduction in computational complexity. However, the inherent drawbacks of using a deep learning model such as large datasets and limited interpretability are the major obstacles in cryptanalysis. In this paper, we introduce another simple data-driven approach using the linear algebraic concept to characterize key-independent balance properties as the kernel of a matrix with empirical parity data. We stack the ciphertext parities obtained under many independent keys into the parity matrix and prove that every mask satisfying the matrix multiplication as zero corresponds exactly to a balance property. Candidates of the balance mask from the test are additionally evaluated by the spurious mask test. We demonstrate the practicality and generality of the kernel methodology on seven lightweight block ciphers spanning SPN with SKINNY, Midori, PRESENT, LED and ARX with SPECK, SIMON, SIMECK. Across these cases, our method recovers known distinguishers and reveals additional non-trivial linear combinations missed by conventional analyses. We additionally position the kernel method relative to other similar methodologies. Our results show that the kernel method provides a rigorous and cipher-agnostic alternative to neural feature exploration and complements division property-based search techniques.

We show that the data deduplication scheme [IEEE Trans. Cloud Comput., 11(1), 156-167, 2023] is flawed due to some inconsistent computations. The scheme tries to propose a hybrid encryption and authentication mechanism based on RSA cryptosystem and pairing cryptosystem, but it has confused the different group operations. In the data sharing phase, the cloud service provider cannot determine which stored tag matches the temporary tag, and fails to return the stored data to the requester. To fix, it should explicitly specify which is randomized by modular exponentiation with an RSA modulus, and which is randomized by point multiplication over the underlying elliptic curve.

This paper presents a new method for approximating the correlations of differential-linear distinguishers. From the perspective of Beyne's geometric approach, the differential-linear correlation is a corresponding coordinate of the \textit{correlation vector} associated with the ciphertext multiset, which can be obtained by using the correlation matrix of the \textit{2-wise form} of the cipher. The composite nature of the correlation matrix leads to a round-based approach to approximate the correlation vector. This simple approximation is remarkably precise, yielding the most accurate estimation for differential-linear correlations in \ascon thus far and the first DL distinguishers for 6-round \ascon-128a initialization. For \present, we present 17-round DL distinguishers, 4 rounds longer than the current record. To apply the round-based approach to ciphers with the large Chi ($\chi$) function as nonlinear functions, we derive theorems to handle the correlation propagation for $\chi$ and its 2-wise form. Strong DL distinguishers for up to 6 rounds of \subterranean and \koala-$p$ are provided, 2 rounds longer than the previous differential and linear distinguishers. Furthermore, the round-based approximation idea is also extended to the higher-order differential-linear distinguishers. We give the first second-order DL distinguisher for 6-round \ascon-128 initialization and the first second-order DL distinguishers for up to 7 rounds of \subterranean and \koala-$p$.

A recent paper by Calderini et al. investigates the use of a CCZ transformation to mask the quadratic central map in a multivariate scheme, providing an instance leading to a system of degree four. A following paper by Caminata et al. presents two methods to reduce the masked system back to a quadratic system. In this work we further study the method based on the quadratic relations between input and output of the masked function, generalizing it and applying to any CCZ transformation (of any quadratic map). Moreover, we study how the existence of these quadratic relations can be used to study whether a function can be CCZ equivalent to a quadratic map and, more generally, to study whether two functions can be CCZ equivalent. In fact, this analysis gives us necessary conditions that can be checked also in relatively large dimensions. Particularly, we completely analyse the cases of vectorial Boolean functions with 1 and 2 output bits.

Solving the Discrete Logarithm problem on the group of points of an elliptic curve is one of the major cryptographic applications of Shor's algorithm. However, current estimates for the number of qubits required remain relatively high, and notably, higher than the best recent estimates for factoring of RSA moduli. For example, recent work by Gidney (arXiv 2025) estimates 2043 logical qubits for breaking 3072-bit RSA, while previous work by Häner et al. (PQCrypto 2020) estimates a requirement of 2124 logical qubits for solving discrete logarithm instances on 256-bit elliptic curves over prime fields. Indeed, for an $n$-bit elliptic curve, the most space-optimized optimized implementation by Proos and Zalka (Quant. Inf. Comput. 2003) gives $5n + o(n)$ qubits, as more additional space is required to store the coordinates of points and compute the addition law. In this paper, we propose an alternative approach to the computation of point multiplication in Shor's algorithm (on input $k$, computing $k P$ where $P$ is a fixed point). Instead of computing the point multiplication explicitly, we use a Residue Number System to compute directly the projective coordinates of $k P$ with low space usage. Then, to avoid performing any modular inversion, we compress the result to a single bit using a Legendre symbol. This strategy allows us to obtain the most space-efficient polynomial-time algorithm for the ECDLP to date, with only $3.12n + o(n)$ qubits, at the expense of an increase in gate count, from $\mathcal{O}(n^3)$ to $\widetilde{\mathcal{O}}(n^4)$. For $n = 256$ we estimate that 1193 qubits would be necessary, with 22 independent runs, using $2^{38.98}$ Toffoli gates each. This represents a much higher gate count than the previous estimate by Häner et al. (roughly $2^{30}$), but half of the corresponding number of qubits (2124).

Non profiled attacks aim to recover secret information from a device without prior knowledge of its leakage model. However, most practical non profiled attacks, such as Differential Power Analysis, Correlation Power Analysis, and Linear Regression-based Attacks (LRA), still depend on a priori leakage assumptions. Designing a generic attack that does not rely on any such assumption therefore remains an open problem and has been actively investigated by the side-channel community for more than a decade. Although Whitnall et al. showed that LRA can be considered generic when all predictors are included, this is not feasible in practice due to inherent multicollinearity issues and the inadequacy of classical distinguishers, which lose discriminating power when targeting injective functions. In this work, we overcome these limitations and propose the first fully generic-compatible non profiled attack. We show that using a Walsh-Hadamard basis enables generic LRA by eliminating multicollinearity and allowing all predictors to be considered without loss of precision. We also introduce new generic-compatible distinguishers tailored to LRA and formally prove their soundness for both linear and non-linear cryptographic operations. Finally, we validate our approach experimentally using simulations and publicly available datasets.

In practice, the amount of side channel leakage related to the random polynomial generation of Dilithium can be limited. In this case, the coe cients of the random polynomial may not be recovered accurately, which then makes the secret key recovery with least square a difficult problem. Therefore, how to recover the secret key used by Dilithium with noisy equations becomes a meaningful concern. In light of this, the ne-grained residual interval screening is proposed to enhance the ability of least square to recover the secret key used by Dilithium. The core idea is to estimate the distributions of the residuals related to both correct and erroneous equations computed with the least square recovered candidate secret key in a pro ling scenario. Then, according to the distribution di erence of the residuals related to two types of equations, an interval can be screened. Note that a majority of erroneous equations are out of the screened interval while a majority of correct equations are in the screened interval. Therefore, least square can be used to recover a more accurate candidate secret key polynomial with equations in the screened interval. By iterating the process several times, the secret key polynomial can be successfully recovered. The e ciency of the proposed technique is veri ed with power traces measured from the Dilithium reference implementation running on a Cortex-M4 processor. The evaluation results show that only several hundreds of power traces are enough to recover the secret key used by Dilithium.

Neural distinguishers for ARX ciphers can exploit information beyond classical difference distributions and several interpretability frameworks have been proposed. In this paper, we study two frameworks for SPECK32/64 by connecting their viewpoints: local constraints of modular addition and Fourier analysis of trained neural distinguishers. We show that the dominant Fourier parities of a raw-pair differential neural distinguisher can be rewritten in the local variables associated with the last modular addition. This representation separates value-dependent variant differential-linear terms from difference-dependent traditional terms, and explains their biases through specific local constraints and branch effects. We further extend the analysis to a boomerang right-quartet setting. We construct a neural distinguisher whose input is only the original ciphertext pair, while positive and negative samples are matched with respect to the observed ciphertext difference. Fourier analysis of this distinguisher reveals dominant value-dependent parities. We trace these terms to borrow synchronization in the first inverse step of the lower boomerang characteristic, yielding specific local conditions. Our results indicate that the dominant Fourier features learned in these settings are observable projections of concrete carry or borrow constraints of the ARX operation.