In Submission 2025/1391 to the IACR Cryptology ePrint Archive, the Inverse Discrete Logarithm Problem (IDLP) is introduced and used to build a key exchange protocol and a KEM. The author claims both classical and post-quantum security for IDLP and therefore for the proposed protocols. It is the purpose of this note to give an efficient quantum algorithm for IDLP, based on the algorithm of Shor. We give an implementation of our algorithm, replacing the use of Shor's algorithm with an oracle.

The Syndrome Decoding Problem (SDP) underpins the security of most code-based cryptographic schemes, and Information Set Decoding (ISD) algorithms are the fastest known solvers for most parameter sets. While ISD is well developed in the binary setting, the landscape for non-binary ISD is less mature. Most $q$-ary methods are straightforward generalizations of their binary counterparts, with the recent projective Stern algorithm being the only exception. However, no existing algorithm is designed to leverage the specific algebraic properties of extension fields. This research gap -- highlighted by the first-round NIST PQC proposal SDitH -- motivates our central question: is decoding over an extension field fundamentally easier than over a prime field of similar size? This work explores whether the algebraic structure of extension fields can accelerate ISD. We analyze several techniques for translating the SDP to the base field, including the expansion map, subfield subcodes, and the trace map. We also develop new BJMM variants that restrict base list vectors to “small” field elements, aiming to counter the performance loss of advanced ISD when $q$ is large. Contrary to our initial intuition, our results provide no evidence of an asymptotic speedup, suggesting that decoding over extension fields is not easier than over prime fields. Additionally, we make two contributions of independent interest: we show that a three-level BJMM algorithm gives a slight improvement over the two-level version for small fields, and we extend Meurer’s proof to show that the complexity of advanced ISD algorithms converges to Prange’s, even when parameters grow simultaneously.

In smart grid (SG), key agreement protocols (KAPs) are used as one of the most prevalent means to establish secure data transmission channels between smart meters (SMs) and service providers (SPs). Quite recently, Wu et al. have indicated the vulnerability of Hu et al.'s KAP to key compromise impersonation (KCI) attack and proposed a security-enhanced one for secure communications of SMs and SPs in SG. Not to undermine the noteworthy contributions of their work, this comment demonstrates that their own KAP, i.e., Wu et al.'s scheme is still vulnerable to KCI attack. Accordingly, we suggest a simple modification to fix the KCI attack issue. Our attack procedure gives some delicate hints to scholars to protect their schemes against the KCI attack in future researches.

Protection of biometric templates is a critical and urgent area of focus. \textbf{IronMask} demonstrates superior recognition performance while protecting facial templates against existing known attacks. In high-level, IronMask can be conceptualized as a fuzzy commitment scheme building on the hypersphere directly. We devise an attack on IronMask targeting on the security notion of renewability. Our attack, termed as \textbf{Probabilistic Linear Regression Attack}, utilizes the linearity of underlying used error correcting code. This attack is the first algorithm to successfully recover the original template when getting multiple protected templates in acceptable time and requirement of storage. We implement experiments on \textbf{IronMask} applied to protect \textbf{ArcFace} that well verify the validity of our attacks. Furthermore, we carry out experiments in noisy environments and confirm that our attacks are still applicable. Finally, we discuss two strategies to mitigate this type of attacks.

Deep learning-based side-channel analysis (DLSCA) has demonstrated remarkable performance over the past few years. Even with limited preprocessing and feature engineering, DLSCA is capable of breaking protected targets, sometimes requiring only a single attack trace. In the DLSCA context, the commonly investigated countermeasures are Boolean masking and desynchronization. While the exact mechanisms of how DLSCA breaks masking are less understood, the core idea behind handling desynchronization is simple. Convolutional neural networks (CNNs) are shift invariant, allowing them to overcome desynchronization. However, considering the importance and practicality of desynchronization countermeasures, we know remarkably little about the limits of CNNs or how to enhance their capabilities when dealing with desynchronization. In this work, we begin with the theoretical foundations of shift and temporal scale equivariance. Afterward, we build a neural network model allowing such equivariance and test it against several commonly considered targets. Our results demonstrate that equivariant CNNs are robust, easy to design, and achieve excellent attack performance. More precisely, we showcase how such a simple model can even outperform recent transformer-based neural networks. Finally, we demonstrate the practical relevance of scale equivariance by showing how an equivariant CNN can learn leakage from a device operating at one clock frequency and generalize to a device with a different clock frequency, a result not previously demonstrated in DLSCA.

In 1994, Langford and Hellman introduced differential-linear (DL) cryptanalysis, with the idea of decomposing the block cipher E into two parts, EU and EL, such that EU exhibits a high-probability differential trail, while EL has a high-correlation linear trail.Combining these trails forms a distinguisher for E, assuming independence between EU and EL. The dependency between the two parts of DL distinguishers remained unaddressed until EUROCRYPT 2019, where Bar-On et al. introduced the DLCT framework, resolving the issue up to one S-box layer. However, extending the DLCT framework to formalize the dependency between the two parts for multiple rounds remained an open problem. In this paper, we first tackle this problem from the perspective of boomerang analysis. By examining the relationships between DLCT, DDT, and LAT, we introduce a set of new tables facilitating the formulation of dependencies between the two parts of the DL distinguisher across multiple rounds. Then, we introduce a highly versatile and easy-to-use automatic tool for exploring DL distinguishers, inspired by automatic tools for boomerang distinguishers. This tool considers the dependency between differential and linear trails across multiple rounds. We apply our tool to various symmetric primitives, and in all applications, we either present the first DL distinguishers or enhance the best-known ones. We achieve successful results against Ascon, AES, SERPENT, PRESENT, SKINNY, TWINE, CLEFIA, WARP, LBlock, Simeck, and KNOT. Furthermore, we demonstrate that, in some cases, DL distinguishers outperform boomerang distinguishers significantly.

In the last few years, the old idea of internal perturbation for multivariate schemes has been resurrected. A form of this method was proposed with application to HFE and UOV and independently by another team for application to Rainbow. Most recently, a newer and more efficient version of internal perturbation was proposed as an enhanced measure for securing HFE for encryption. This efficient method, known as the LL' construction, is designed to add little complexity to HFE decryption while increasing the rank of the resulting map to resist the now very effective cryptanalyses powered by MinRank. The basic idea of the construction is to have two small lists of binary linear forms which when multiplied produce rank $1$ quadratic forms. Random linear combinations of these products are then added to each of the HFE equations, resulting in a masked HFE. The main trick to make the scheme usable is to encrypt an send many random messages so that statistically it is likely that the legitimate user can find a ciphertext that is not perturbed by the construction and which may be decrypted as a plain HFE ciphertext. We show that this approach is not secure. In particular, we present a method to recover the noise support, a collection of quadratic forms spanning the set of LL' quadratic forms. We then are able to filter out the effect of these maps to recover a compatible HFE map. Finally, we are able to complete the key recovery, achieving efficiently an equivalent private key.

In 2023, Barreto and Zanon proposed a three-round Schnorr-like blind signature scheme, leveraging zero-knowledge proofs to produce one-time signatures as an intermediate step of the protocol. The resulting scheme, called BZ, is proven secure in the discrete-logarithm setting under the one-more discrete logarithm assumption with (allegedly) resistance to the Random inhomogeneities in a Overdetermined Solvable system of linear equations modulo a prime number $p$ attack, commonly referred to as ROS attack. The authors argue that the scheme is resistant against a ROS-based attack by building an adversary whose success depends on extracting the discrete logarithm of the intermediate signing key. In this paper, however, we describe a distinct ROS attack on the BZ scheme, in which a probabilistic polynomial-time attacker can bypass the zero-knowledge proof step to break the one-more unforgeability of the scheme. We also built a BZ variant that, by using one secure hash function instead of two, can prevent this particular attack. Unfortunately, though, we show yet another ROS attack that leverages the BZ scheme's structure to break the one-more unforgeability principle again, thus revealing that this variant is also vulnerable. These results indicate that, like other Schnorr-based strategies, it is hard to build a secure blind signature scheme using BZ's underlying structure.

HPPC is a multivariate signature scheme submitted to the NIST PQC standardization process in response to the recent call for additional signature schemes. We show that, despite some non-standard notational choices in the submission document, HPPC can be viewed as a special case of the well-studied, but broken for all practical parameters, HFE signature scheme. We further show that the HPPC construction introduces additional structure that further weakens the scheme. For instance, the central map has $Q$-rank $2$ independently from the degree $D$ of the central polynomial that is used. Using these observations, we show that HPPC is weaker against the direct attack than claimed in the submission document and more crucially that all parameter sets can be practically broken using MinRank techniques. For instance, with a very naive implementation, we have been able to recover an equivalent key in approximately 8 minutes for security level 2, an hour and a half for security level 4, and slightly more than 7 hours for security level 5.

This paper studies the Twinkle family of low-latency symmetric key schemes designed by Wang et al. (CiC 2024). In particular, it presents cryptanalysis of both the mode and the underlying primitive. Twinkle is a PRF-based design, and an authenticated encryption scheme Twinkle-AE is specified based on a dedicated PRF called Twinkle-PRF. To achieve low latency, Twinkle-PRF uses a large key and state to produce sufficient randomness in a single step. Twinkle-AE uses a 1024- or 512-bit key for authentication and generates a $t$-bit tag, where $t \in \{64, 128\}$. It claims to provide $t$ bits of integrity. Several Twinkle-AE parameter sets claim higher confidentiality than integrity. In this setup, for any ciphertext, an adversary can obtain the message after $O(2^t)$ decryption attempts by guessing the tag, allowing attacks in the chosen-ciphertext setting. We show that a 1024- or 512-bit authentication key can be recovered using only $O(2^t)$ queries. The recovered authentication key enables the generation of valid ciphertexts for arbitrary plaintexts, thus achieving universal forgery. In the second part of the paper, we perform cryptanalysis on reduced-round variants of the 1280-bit public permutation Twinkle-P, which serves as a core component of Twinkle-PRF. We investigate impossible differential, zero-correlation linear, integral, and differential-linear distinguishers by developing automated analytic tools. We provide practical distinguishers for up to 5 rounds, and the longest distinguisher reaches 6 rounds with a complexity of $2^{74.32}$. This surpasses the round bounds evaluated by the designers. We stress that our attacks on mode exploits the gap between the claimed confidentiality and integrity levels, thus have no impact on the parameter sets having the same security level. Our attacks on the permutation do not have any significant impact on the whole specifications. Moreover, we note that Twinkle-AE-512b/Twinkle-AE-1024b and Twinkle-PA remain secure, and the versions we attacked would also be secure if the claimed confidentiality level matched the integrity level.

Deep learning (DL)-based side-channel analysis (SCA) has emerged as a powerful approach for extracting secret information from cryptographic devices. However, its performance often deteriorates when targeting implementations protected by masking and desynchronization-based countermeasures, or when analyzing long side-channel traces. In earlier work, we proposed EstraNet, a Transformer Network (TN)-based model designed to address these challenges by capturing long-distance dependencies and incorporating shift-invariant attention mechanisms. In this work, we perform an in-depth analysis of the internal behavior of EstraNet and propose methods to further enhance its effectiveness. First, we introduce {\bf DL-ProVe} (Deep Learning Leakage Propagation Vector Visualization), a novel technique for visualizing how leakage from secret shares in a masked implementation propagates and recombines into the unmasked secret through the layers of a DL model trained for SCA. We then apply DL-ProVe to EstraNet, providing the first detailed explanation of how leakage is accumulated and combined within such an architecture. Our analysis reveals a critical limitation in EstraNet’s multi-head GaussiP attention mechanism when applied to long traces. Based on this insights, we identify a new architectural hyperparameter which enables fine-grained control over the initialization of the attention heads. Our experimental results demonstrate that tuning this hyperparameter significantly improves EstraNet’s performance on long traces (with upto 250K features), allowing it to reach the guessing entropy 1 using only 3 attack traces while the original EstraNet model fails to do so even with 5K traces. These findings not only deepen our understanding of EstraNet’s internal workings but also introduce a robust methodology for interpreting, diagnosing, and improving DL models for SCA.

The cube attack is one of the most powerful attacks on stream ciphers, with recovering the superpoly as its key step. The core monomial prediction is the state-of-the-art technique for superpoly recovery, which can reach 851 rounds for Trivium thus far (EUROCRYPT 2024). The core monomial prediction heavily relies on the trail enumeration which is the bottleneck for its efficiency. This paper further explores the potential of the core monomial prediction for Trivium by constructing a composite representation for the superpoly. This representation allows us to detect the algebraic structure of the superpoly under specific conditions on the intermediate variables, without the computational burden of trail enumerations. Leveraging these discovered conditions, we successfully recovered weak-key superpolies for 852-round Trivium, establishing the first cryptanalytic result against 852-round Trivium in the literature to date.

We consider the multivariate scheme $\texttt{Pesto}$, which was introduced by Calderini, Caminata, and Villa. In this scheme, the public polynomials are obtained by applying a CCZ transformation to a set of quadratic secret polynomials. As a consequence, the public key consists of polynomials of degree $4$. In this work, we show that the public degree $4$ polynomial system can be efficiently reduced to a system of quadratic polynomials. This seems to suggest that the CCZ transformation may not offer a significant increase in security, contrary to what was initially believed.

We provide minimal counterexamples for the security of the BitVM3 garbling scheme: our attack allows the evaluator to forge input and output wires. Then we use the same idea to exhibit an attack on the forward label propagation garbling scheme proposed in a more recent paper. In both cases, the authenticity property of the garbling scheme is broken.

This tutorial provides a practical introduction to Deep Learning-based Side-Channel Analysis (DLSCA), a powerful approach for evaluating the security of cryptographic implementations. Leveraging publicly available datasets and a Google Colab service, we guide readers through the fundamental steps of DLSCA, offering clear explanations and code snippets. We focus on the core DLSCA framework, providing references for more advanced techniques, and address the growing interest in this field driven by emerging standardization efforts like AIS 46. This tutorial is designed to be accessible to researchers, students, and practitioners seeking to learn practical DLSCA techniques and improve the security of cryptographic systems.

This paper presents a novel single-trace side-channel attack on FALCON, a lattice-based post-quantum digital signature protocol recently approved for standardization by NIST. We target the discrete Gaussian sampling operation within FALCON’s key generation scheme and demonstrate that a single power trace is sufficient to mount a successful attack. Notably, negating the results of a 63-bit right-shift operation on 64-bit secret values leaks critical information about the assignment of ‘-1’ versus ‘0’ to intermediate coefficients during sampling. These leaks enable full recovery of the secret key. We demonstrate a ground-up approach to the attack on an ARM Cortex-M4 microcontroller executing both the reference and optimized implementations from FALCON’s NIST round 3 software package. We successfully recovered all of the secret polynomials in FALCON. We further quantify the attacker’s success rate using a univariate Gaussian template model, providing generalizable guarantees. Statistical analysis with over 500,000 tests reveals a per-coefficient success rate of 99.9999999478% and a full-key recovery rate of 99.99994654% for FALCON-512. We verify that this vulnerability is present in all implementations included in FALCON’s NIST submission package. This highlights the vulnerability of current software implementations to single-trace attacks and underscores the urgent need for single-trace- resilient software in embedded systems.

Side-channel analysis (SCA) has posed a significant threat to systems for nearly three decades. Numerous practical demonstrations have targeted everyday devices, such as smart cards, cryptocurrency wallets, and smartphones. However, much of the research in the public domain has focused on low-end microcontrollers, limiting our understanding of the challenges involved in attacking more complex systems. In this work, we conduct a reality check on SCA by targeting a high-performance ARM Cortex-A72 out-of-order processor, commonly found in smartphones. We evaluate the practical effort required for key recovery attacks, considering various threat models, from basic to advanced. Our results show that while basic approaches fail, advanced approaches like deep learning-based SCA can successfully recover the secret key. This multi-tier evaluation approach is crucial for comprehensive risk assessment and informed decision-making regarding mitigation strategies, balancing security, performance, and area constraints.

We propose a novel partial key exposure attack on common prime RSA by leveraging lattice-based techniques. In common prime RSA, the primes $p$ and $q$ are defined as $p=2ga+1$ and $q=2gb+1$ for a common prime $g$. Recently, Zheng introduced the first partial key exposure attack on this scheme; however, it is limited to instances where $g > N^{1/4}$. In contrast, our work investigates deeper into partial key exposure attacks by presenting a unified generic case that targets one consecutive unknown block of the private key. By employing a lattice-based solving strategy for trivariate integer polynomials, we can effectively identify additional weak private keys that are vulnerable to partial exposure. Extensive numerical experiments validate the correctness and practicality of our proposed attack on common prime RSA.

In this paper, we present a new small private exponent attack on RSA by combining continued fractions and Coppersmith's techniques. Our results improve upon previous bounds, including Herrmann-May's attack, by leveraging a crucial relation derived from continued fraction. Additionally, we extend the range of vulnerable small private exponents by considering the partial leakage of prime factors or their sum. Our main result establishes an improved attack bound $ d < N^{1-\alpha/3-\gamma/2} $, where $ \alpha := \log_{N} e $ and $ \gamma := \log_{N} |p+q-S| $, with $ S $ being an approximation of the prime sum $ p+q $. Furthermore, we explore more applications of our main attack in scenarios where the primes share some most or least significant bits. The validity of our proposed main attack is confirmed through numerical experiments, demonstrating its improved performance over existing attacks.

Recently, a more efficient attack on the initial tropical Stickel protocol has been proposed, different from the previously known Kotov-Ushakov attack, yet equally guaranteed to succeed. Given that the Stickel protocol can be implemented in various ways, such as utilizing platforms beyond the tropical semiring or employing alternative commutative matrix ``classes'' instead of polynomials, we firstly explore the generalizability of this new attack across different implementations of the Stickel protocol. We then conduct a comprehensive security analysis of the Stickel protocol based on Linde-de la Puente (LdlP) matrices. Additionally, we extend the concept of LdlP matrices beyond the tropical semiring, generalizing it to a broader class of semirings and include a discussion of the centralizer of such matrices over the tropical semiring.

HQC is a code-based cryptosystem that has recently been announced for standardization after the fourth round of the NIST post-quantum cryptography standardization process. During this process, the NIST specifically required submitters to provide two kinds of implementation: a reference one, meant to serve lisibility and compliance with the specifications; and an optimized one, aimed at showing the performance of the scheme alongside other desirable properties such as resilience against implementation misuse or side-channel analysis. While most side-channel attacks regarding PQC candidates running in this process were mounted over reference implementations, very few consider the optimized, allegedly side-channel resistant (at least, constant-time), implementations. Unfortunately, HQC optimized version only targets x86-64 with Single Instruction Multiple Data (SIMD) support, which reduces the code portability, especially for non-generalist computers. In this work, we present two power side-channel attacks on the optimized HQC implementation with just the SIMD support deactivated. We show that the power leaks enough information to recover the private key, assuming the adversary can ask the target to replay a legitimate decryption with the same inputs. Under this assumption, we first present a key-recovery attack targeting standard Instruction Set Architectures (ARM T32, RISC-V, x86-64) and compiler optimization levels. It is based on the well known Hamming Distance model of power consumption leakage, and exposes the key from a single oracle call. During execution on a real target, we show that a different leakage, stemming from to the micro-architecture, simplifies the recovery of the private key. This more direct second attack, succeeds with a 99% chance from 83 executions of the same legitimate decryption. While the weakness leveraged in this work seems quite devastating, we discuss simple yet effective and efficient countermeasures to prevent such a key-recovery.

Side-channel attacks (SCA) pose a significant threat to cryptographic implementations, including those designed to withstand the computational power of quantum computers. This paper introduces the first side-channel attack on an industry-grade post-quantum cryptography implementation. Specifically, we present a Correlation Power Analysis (CPA) attack targeting the open-source hardware implementation of ML-DSA within a Silicon Root of Trust framework developed through a multi-party collaboration involving leading technology companies. Our attack focuses on the modular reduction process that follows the Number Theoretic Transform-based polynomial pointwise multiplication. By exploiting side-channel leakage from a distinctive unique reduction algorithm and leveraging the zeroization mechanism used to securely erase sensitive information by clearing internal registers, we significantly enhance the attack's efficacy. Our findings reveal that an adversary can extract the secret keys using only 10,000 power traces. With access to these keys, an attacker could forge signatures for certificate generation, thereby compromising the integrity of the root of trust. This work highlights the vulnerabilities of industry-standard root-of-trust systems to side-channel attacks. It underscores the urgent need for robust countermeasures to secure commercially deployed systems against such threats.

Many protocols, like HTTP, FTP, POP3, and SMTP, were origi- nally designed as synchronous plaintext protocols – commands and data are sent in the clear, and the client waits for the response to a pending request before sending the next one. Later, two main solutions were introduced to retrofit these protocols with TLS protection. (1) Implicit TLS: Designate a new, well-known TCP port for each protocol-over-TLS, and start with TLS immediately. (2) Opportunistic TLS: Keep the original well-known port and start with the plaintext protocol, then switch to TLS in response to a command like STARTTLS. In this work, we present a novel weakness in the way TLS is integrated into popular application layer protocols through implicit and opportunistic TLS. This weakness breaks authentication, even in modern TLS implementations if both implicit TLS and oppor- tunistic TLS are supported at the same time. This authentication flaw can then be utilized to influence the exchanged messages after the TLS handshake from a pure MitM position.In contrast to previ- ous attacks on opportunistic TLS, this attack class does not rely on bugs in the implementations and only requires one of the peers to support opportunistic TLS. We analyze popular application layer protocols that support opportunistic TLS regarding their vulnerability to the attack. To demonstrate the practical impact of the attack, we analyze exploita- tion techniques for HTTP (RFC 2817) in detail, and show four different exploit directions. To estimate the impact of the attack on deployed servers, we conducted a series of IPv4-wide scans over multiple protocols and ports to check for support of opportunistic TLS. We found that support for opportunistic TLS is still widespread for many application protocols, with over 3 million servers support- ing both, implicit and opportunistic TLS at the same time. In the case of HTTP, we found 20,121 servers that support opportunistic HTTP across 35 ports, with 2,268 of these servers also supporting HTTPS and 539 using the same domain names for implicit HTTPS, presenting an exploitable scenario.

In this work, we present methods for conducting higher-order non-profiled side-channel attacks on Lattice-Based Cryptography (LBC). Our analysis covers two scenarios: one where the device leakage is known and follows Hamming weight model, and another where the leakage model is not Hamming weight based and unknown to the attacker. We focus on the Post-Quantum Cryptography (PQC) standards, the Dilithium digital signature (i.e. ML-DSA) and the Kyber key encapsulation (i.e. ML-KEM) algorithms. For Hamming weight leakage, we develop efficient higher-order Correlation Power Analysis (HOCPA) attacks in which the attacker must compute a function known as the optimal prediction function. We revisit the definition of optimal prediction function and introduce a recursive method for computing it efficiently. Our approach is particularly useful when a closed-form formula is unavailable, as in LBC. Then, we introduce sin and cos prediction functions, which prove optimal for HOCPA attacks against second and higher-order masking protection. We validate our methods through simulations and real-device experiments on open-source masked implementations of Dilithium and Kyber on an Arm Cortex-M4. we achieve full secret-key recovery using only 700 and 2400 traces for second and third-order masked implementations of Dilithium, and 2200 and 14500 traces for second and third-order masked implementations of Kyber, respectively. For the unknown leakage scenarios, we leverage generic Side-Channel Analysis (SCA) distinguishers. A key challenge here is the injectivity of modular multiplications in NTT based polynomial multiplication, typically addressed by bit-dropping in the literature. However, we experimentally show that bit-dropping is largely inefficient against protected implementations of LBC. To overcome this limitation, we present a novel two-step attack to Kyber, combining generic distinguishers and lattice reduction techniques. Our approach decreases the number of predictions from q^2 to q and does not rely on bit-dropping. Our experimental results demonstrate a speed-up of up to 23490x in attack run-time over the baseline along with improved success rate.

Side-channel analysis (SCA) exploits the dependency of a cryptographic device's power consumption or electromagnetic (EM) emissions on the data being processed to extract secret keys. While deep learning (DL) has advanced SCA, most existing models struggle to scale to long traces and tend to be effective only against specific countermeasures. Therefore, their applicability is limited in implementations that employ multiple countermeasures simultaneously. This paper introduces UNIDLE, a UNIfied framework for Deep LEarning-based SCA, for building scalable and robust models for long traces. UNIDLE employs a hierarchical divide-and-conquer strategy: it segments each trace into smaller parts, which are independently processed by base models, and then aggregates their outputs using a top-level model. This approach enables efficient information extraction from long traces while making the model robust against combinations of masking and desynchronization-based countermeasures. UNIDLE also naturally supports multi-task learning, a useful DL approach for attacking countermeasures such as shuffling, while being less susceptible to the capacity bottlenecks observed in existing multi-task models. Leveraging this framework, we develop HierNet, a novel DL model that integrates a shift-invariant base model and a transformer-based top-level model. HierNet demonstrates strong empirical performance across three datasets of masked implementations, reaching the guessing entropy 1 with fewer than 50 attack traces--where several state-of-the-art benchmark models fail even with up to 5K attack traces. Experiments on traces protected by shuffling show that HierNet can reach the guessing entropy 1 using 95% fewer attack traces than a state-of-the-art multi-task benchmark while requiring only one-tenth of the model size.

This paper introduces the weighted sum correlation analysis method, a profiled higher-order side-channel attack that quantifies the significance of time-domain samples based on a chosen leakage model. We also demonstrate the utility of the Hilbert transform in side-channel analysis, showing how its phase-shifting property can be exploited to construct an effective fused score that combines multiple correlation coefficients into a single metric. We validate the presented method on the challenging case of the AES-CCM accelerator in a commercial Bluetooth chip, leveraging RF signals captured via a software-defined radio as a side channel. Compared to the correlation analysis methods presented at RWC'25 and CHES'25, the weighted sum approach achieves at least a threefold reduction in the number of traces required for key recovery. Remarkably, it also outperforms deep learning-based analysis.

Differential meet-in-the-middle (MITM) cryptanalysis, recently introduced by Boura et al., has emerged as a powerful and versatile technique for assessing the security of modern block cipher designs. Since its introduction, this method has been effectively applied to a variety of block ciphers, including different variants of SKINNY, CRAFT, and AES. However, identifying such attacks manually–especially on bit-oriented ciphers with large block sizes–can be a complex and error-prone process, which underscores the growing importance of automated solutions in this domain. In this work, we present, for the first time to the best of our knowledge, a novel and efficient automated tool for constructing optimized differential MITM attacks on bit-oriented block ciphers, with a particular focus on AndRX designs. Our framework begins by modeling an efficient constraint programming (CP) model to search for single-key optimal differential trails in AndRX ciphers. Building on this, we propose a unified bitwise CP model to automatically construct optimized differential MITM attacks within the same design framework. Furthermore, we incorporate two dedicated optimization strategies–namely, the equivalent subkey technique and the selective key guessing technique–both of which are tailored to the structural properties of AndRX ciphers and significantly enhance key recovery efficiency. Additionally, we also apply two additional optimization techniques: the parallel partitioning technique and the reducing data with imposed conditions techniques to further enhance the differential MITM attack on AndRX ciphers. To demonstrate the practical effectiveness of our tool, we apply it to all versions of SIMON and Simeck, two widely studied representatives of the AndRX family, and report improved cryptanalytic results. Specifically, we present differential MITM attacks on SIMON-32-64, SIMON-48-96, SIMON-64-128, and SIMON-96-144, covering 23, 25, 32, and 38 rounds, respectively. All of these results represent improvements in the number of attacked rounds compared to the best known differential attacks, classical meet-in-the-middle (MITM), and Demirci-Selçuk MITM (DS-MITM) attacks on the corresponding versions of SIMON. For instance, we present a 37-round differential MITM attack on SIMON-96-144, which extends the best known differential, classical MITM, and DS-MITM attacks by 1, 17, and 18 rounds, respectively. In the case of Simeck, we report a 29-round differential MITM attack on Simeck-48-96, improving the previous best differential attack by one round. These results demonstrate the strength and versatility of our automated tool. Importantly, our automated method for constructing differential MITM attacks operates at the bit level and is generic in nature, making it applicable to a broad class of bit-oriented block ciphers beyond the AndRX family.

Statistical Ineffective Fault Attack (SIFA) presents a critical threat to cryptographic implementations by circumventing conventional detection-based countermeasures effective against traditional fault attacks. Particularly, SIFA operates via two mechanisms: SIFA-1 exploits fault effectiveness dependency on target values, while SIFA-2 leverages conditional propagation of faulted values based on sensitive intermediates. Recent studies suggest that, masking, mainly a side-channel protection, also exhibits promising resistance to SIFA-1, such as prime masking. In this paper, we systematically evaluate the resilience of Inner Product Masking (IPM) against SIFA, which has been established in prior works as a powerful side-channel-resistant alternative to Boolean masking. Specifically, with regard to SIFA-1, our theoretical analysis demonstrates that Inner Product (IP) encoding provides stronger SIFA-1 resistance than both Boolean and prime masking under generic multi-bit fault models using various fault types. More interestingly, an equivalence between Side-channel and SIFA-1 security has been theoretically established for IP encoding, indicating that optimal IP encoding exists that simultaneously provides the highest side-channel resistance and maximizes the complexity of effective SIFA-1 attacks. For SIFA-2, our analysis reveals that IPM’s protection remains fundamentally bounded by the computational field size, consistent with previous results in this regard, e.g., for prime field masking. However, some vulnerabilities to persistent faults are anticipated for the most recently proposed IPM multiplication gadget. Given the compatibility with existing ciphers and demonstrated superior resistance against SIFA-1, IPM emerges as a more promising fault-resistant encoding technique compared to prime masking.

In the present work we address the robustness of the Polynomial Learning With Errors problem extending previous results in Blanco-Chacón et al. and in Elias et al. In particular, we produce two kinds of new distinguishing attacks: a) we generalize Blanco-Chacón et al. to the case where the defining polynomial has a root of degree up to 4, and b) we widen and refine the most general attack in Elias et al. to the non-split case and determine further dangerous instances previously not detected. Finally, we exploit our results in order to show vulnerabilities of some cryptographically relevant polynomials.

Fully Homomorphic Encryption (FHE) inherently lacks data integrity mechanisms, allowing a malicious server to arbitrarily tamper with the data associated with FHE computations on its end. A series of works named reaction attacks further demonstrates that a malicious server can exploit this lack of integrity checks to carry out interactive full-key recovery attacks on state-of-the-art FHE schemes. In this paper, we propose an efficient solution to this problem by utilizing the concept of the Merkle tree. Our solution uses cryptographic hash functions to ensure the integrity of data involved in FHE computations. The efficiency of our solution comes from the lower sizes and ease of computations of the hash values, which subsequently leads to a reduction in both the network and computation overhead. Given that protecting the entire FHE circuit can lead to tremendous network overhead, we further perform scheme-specific optimizations by identifying a small portion of the FHE circuit, protecting which is sufficient to thwart these reaction-based attacks. Finally, we propose a framework that evaluates different FHE schemes based on the overhead that will be incurred when protecting these schemes through our solution. Our framework can be leveraged by application developers to choose an optimal FHE scheme in terms of both performance and overhead.

In this note we study the proposed cipher Synergy and describe a full round differential with probability $2^{-21.29}$. The claims have been experimentally verified.

Censorship resilience is a fundamental assumption underlying the security of blockchain protocols. Additionally, the analysis of blockchain security from an economic and game theoretic perspective has been growing in popularity in recent years. In this work, we present a surprising rational censorship attack on blockchain censorship resilience when we adopt the analysis of blockchain security from a game theoretic lens and assume all users are rational. In our attack, a colluding group with sufficient voting power censors the remainder nodes such that the group alone can gain all the rewards from maintaining the blockchain. We show that if nodes are rational, coordinating this attack just requires a public read and write blackboard and we formally model the attack using a game theoretic framework. Furthermore, we note that to ensure the success of the attack, nodes need to know the total true voting power held by the colluding group. We prove that the strategy to join the rational censorship attack and also for nodes to honestly declare their power is a subgame perfect equilibrium in the corresponding extensive form game induced by our attack. Finally, we discuss the implications of the attack on blockchain users and protocol designers as well as some potential countermeasures.

Classic McEliece is one of the code-based Key Encapsulation Mechanism finalists in the ongoing NIST post-quantum cryptography standardization process. Several key-recovery side-channel attacks on the decapsulation algorithm have already been published. However none of them discusses the feasibility and/or efficiency of the attack in the case of noisy side-channel acquisitions. In this paper, we address this issue by proposing two improvements on the recent key-recovery attack published by Drăgoi et al.. First, we introduce an error correction algorithm for the lists of Hamming weights obtained by side-channel measurements, based on the assumption, validated experimentally, that the error on a recovered Hamming weight is bounded to $\pm1$. We then offer a comparison between two decoding efficiency metrics, the theoretical minimal error correction capability and an empirical average correction probability. We show that the minimal error correction capability, widely used for linear codes, is not suitable for the (non-linear) code formed by the lists of Hamming weights. Conversely, experimental results show that out of 1 million random erroneous lists of $2t=128$ Hamming weights, only 2 could not be corrected by the proposed algorithm. This shows that the probability of successfully decoding a list of erroneous Hamming weights is very high, regardless of the error weight. In addition to this algorithm, we describe how the secret Goppa polynomial $g$, recovered during the first step of the attack, can be exploited to reduce both the time and space complexity of recovering the secret permuted support $\mathcal{L}$.

Coppersmith's method is a well-known and practical method for solving polynomial modular equations involved in some cryptosystems such as RSA. An important and tedious task in this method consists in computing the asymptotic bounds. In this work, we address the challenge of computing such asymptotic bounds by introducing the Sumsets theory from Additive Combinatorics as a new analytical tool, which significantly streamlines manual calculations. More precisely, we develop the first provable algorithm for determining these asymptotic bounds, whereas the recent methods based on simple Lagrange interpolation are heuristic. Moreover, the experiments showed that our method is much more efficient than the previous method in practice. We also employ our method to improve the cryptanalytic results for the Commutative Isogeny Hidden Number Problem. Our approach may deepen the understanding of Coppersmith's method and inspire more security analysis methodologies.

Post-Quantum Cryptography (PQC) algorithms should remain secure even in the presence of quantum computers. Although the security of such schemes is guaranteed at the algorithmic level, real-world implementations often suffer from other vulnerabilities like Side-Channel Attacks (SCA). This Systematization of Knowledge (SoK) paper investigates side-channel attacks targeting implementations of PQC algorithms. This work categorizes attacks from an adversarial perspective to identify the most vulnerable components of the algorithms' implementations and highlights unexplored parts in current implementations. In addition, it reviews and analyzes the efficiency and efficacy of existing countermeasures to SCA in current hardware implementations. This approach helps identify countermeasures that provide broader protection and highlights characteristics needed for future secure implementations. Our findings offer guidance in strengthening existing systems and developing more efficient defenses against side-channel attacks.

Simple power analysis (SPA) attacks and their extensions, profiled and soft-analytical side-channel attacks (SASCA), represent a significant threat to the security of cryptographic devices and remain among the most powerful classes of passive side-channel attacks. In this work, we analyze how numeric representations of secrets can affect the amount of exploitable information leakage available to the adversary. We present an analysis of how mutual information changes as a result of the integer ring size relative to the machine word-size. Furthermore, we study the Redundant Number Representation (RNR) countermeasure and show that its application to ML-KEM can resist the most powerful SASCA attacks and provides a low-cost alternative to shuffling. We eval- uate the performance of RNR-ML-KEM with both simulated and prac- tical SASCA experiments on the ARM Cortex-M4 based on a worst-case attack methodology. We show that RNR-ML-KEM sufficiently renders these attacks ineffective. Finally, we evaluate the performance of the RNR-ML-KEM NTT and INTT and show that SPA security can be achieved with a 62.8% overhead for the NTT and 0% overhead for the INTT relative to the ARM Cortex-M4 reference implementation used.

Deep learning-based side-channel attack (DLSCA) has become the dominant paradigm for extracting sensitive information from hardware implementations due to its ability to learn discriminative features directly from raw side-channel traces. A common design choice in DLSCA involves embedding traces in Euclidean space, where the underlying geometry supports conventional objectives such as classification or contrastive learning. However, Euclidean space is fundamentally limited in capturing the multi-level hierarchical structure of side-channel traces, which often exhibit both coarse-grained clustering patterns (e.g., Hamming weight similarities) and fine-grained distinctions (e.g., instruction-level variations). These limitations adversely affect the discriminability and generalization of learned representations, particularly across diverse datasets and leakage models. In this work, we propose HypSCA, a dual-space representation learning method that embeds traces in hyperbolic space to exploit its natural ability to model hierarchical relationships through exponential volume growth. In contrast to existing approaches, HypSCA jointly combines hyperbolic structure modeling with local discriminative learning in Euclidean space, enabling the preservation of global hierarchies while enhancing fine-grained feature separation. Extensive experiments on multiple public datasets demonstrate that HypSCA achieves up to 51.6% improvement in attack performance over state-of-the-art DLSCA methods, consistently enhancing generalization across diverse datasets and leakage models.

This paper introduces the Pad Thai message recovery attack on REDOG, a rank-metric code-based encryption scheme selected for the second round of evaluation in the Korean Post-Quantum Cryptography (KPQC) competition. The attack exploits the low rank weight of a portion of the ciphertext to construct multiple systems of linear equations, one of which is noise-free and can be solved to recover the secret message. The Pad Thai attack significantly undermines the security of REDOG, revealing that its provided security is much lower than originally claimed.

We propose a dimension-reducing transformation on Group Ring Learning with Errors (GRLWE) samples. We exhibit an efficiently computable isomorphism which takes samples defined over the group rings used in the construction of GRLWE to twice as many samples defined over matrix rings, in half the dimension. This is done by composing two maps: the first map is a transformation showing that the group rings used are orders in central simple algebras, and the second map takes the obtained central simple algebra to a matrix ring. When combined with lattice reduction on the resulting matrix samples, this gives an attack on the GRLWE problem. We extend this attack to other groups proposed for cryptographic use by the creators of GRLWE, and display some numerical results quantifying the effects of the transformation, using the `Lattice Estimator'. We then give a family of groups from which GRLWE-style group rings can be constructed which are immune to our attack, namely the generalized quaternion groups. Finally, we discuss the merits and vulnerabilities of a number of different forms of structured LWE.

At CRYPTO 2020, Dachman-Soled et al. introduced a framework for to analyze the security loss of Learning with Errors ($\text{LWE}$), which enables the incremental integration of leaked hints into lattice-based attacks. Later Nowakowski and May at ASIACRYPT 2023 proposed a novel method capable of integrating and combining an arbitrary number of both perfect and modular hints for the LWE secret within a unified framework, which achieves better efficiency in constructing the lattice basis and makes the attacks more practical. In this paper, we first consider solving LWE with independent hints about both the secret and errors. Firstly, we introduce a novel approach to embed the hints for secret into the $\text{LWE}$ lattice by just matrix multiplication instead of the LLL reduction as in Nowakowski and May's attack, which further reduces the time complexity to construct the lattice basis. For example, given 234 perfect hints about CRYSTALS-KYBER 512, our method reduces the running time from 2.16 hours to 0.35 hours. Secondly, we show how to embed the hints about errors into the obtained lattice basis.

In the last years, Deep Learning algorithms have been browsed and applied to Side-Channel Analysis in order to enhance attack’s performances. In some cases, the proposals came without an indepth analysis allowing to understand the tool, its applicability scenarios, its limitations and the advantages it brings with respect to classical statistical tools. As an example, a study presented at CHES 2021 proposed a corrective iterative framework to perform an unsupervised attack which achieves a 100% key bits recovery. In this paper we analyze the iterative framework and the datasets it was applied onto. The analysis suggests a much easier and interpretable way to both implement such an iterative framework and perform the attack using more conventional solutions, without affecting the attack’s performances.

The benefits of using Deep Learning techniques to enhance side-channel attacks performances have been demonstrated over recent years. Most of the work carried out since then focuses on discriminative models. However, one of their major limitations is the lack of theoretical results. Indeed, this lack of theoretical results, especially concerning the choice of neural network architecture to consider or the loss to prioritize to build an optimal model, can be problematic for both attackers and evaluators. Recently, Zaid et al. addressed this problem by proposing a generative model that bridges conventional profiled attacks and deep learning techniques, thus providing a model that is both explicable and interpretable. Nevertheless the proposed model has several limitations. Indeed, the architecture is too complex, higher-order attacks cannot be mounted and desynchronization is not handled by this model. In this paper, we address the first limitation namely the architecture complexity, as without a simpler model, the other limitations cannot be treated properly. To do so, we propose a new generative model that relies on solid theoretical results. This model is based on conditional variational autoencoder and converges towards the optimal statistical model i.e. it performs an optimal attack. By building on and extending the state-of-the-art theoretical works on dimensionality reduction, we integrate into this neural network an optimal dimensionality reduction i.e. a dimensionality reduction that is achieved without any loss of information. This results in a gain of $\mathcal{O}(D)$, with $D$ the dimension of traces, compared to Zaid et al. neural network in terms of architecture complexity, while at the same time enhancing the explainability and interpretability. In addition, we propose a new attack strategy based on our neural network, which reduces the attack complexity of generative models from $\mathcal{O}(N)$ to $\mathcal{O}(1)$, with $N$ the number of generated traces. We validate all our theoretical results experimentally using extensive simulations and various publicly available datasets covering symmetric, asymmetric pre and post-quantum cryptography implementations.

Information Set Decoding (ISD) algorithms currently offer the most powerful tool to solve the two archetypal problems of coding theory, namely the codeword finding roblem and the syndrome decoding problem. Traditionally, ISD have primarily been studied for linear codes over finite fields, equipped with the Hamming metric. However, recently, other possibilities have also been explored. These algorithms have been adapted to different ambient spaces and metrics, such as the rank metric or the Lee metric over $\mathbb{Z}_m$. In this paper, we propose a general framework for decoding ring-linear codes that exploits the underlying ring structure to improve traditional approaches. The core idea is to project the decoding instance onto a smaller alphabet, which may enable more efficient decoding algorithms. Our framework is compatible with any coordinate-additive metric, including Hamming and Lee, and can also be adapted to the rank metric. However, we show that the effectiveness of this approach strongly depends on the chosen metric. We illustrate how this framework can be leveraged to design decoding algorithms for the two aforementioned problems in Hamming, rank, and Lee metrics, along with their range of effectiveness. For each case, we provide the average computational complexity of the resulting algorithms.

The module-Lattice Isomorphism Problem (module-LIP) was introduced by Ducas et al. (ASIACRYPT 22) in~\cite{HAWK:cryptoeprint:2022/1155}, and used within the signature scheme and NIST candidate HAWK. In~\cite{modLIPtotallyreal}, Mureau et al. (EUROCRYPT24) pointed out that over certain number fields $F$, the problem can be reduced to enumerating solutions of $x^2 + y^2 = q$ (where $q \in \O_F$ is given and $x,y \in \O_F$ are the unknowns). Moreover one can always reduce to a similar equation which has only \textit{few} solutions. This key insight led to a heuristic polynomial-time algorithm for solving module-LIP on those specific instances. Yet this result doesn't threaten HAWK for which the problem can be reduced to enumerating solutions of $x^2 + y^2 + z^2 + t^2 = q$ (where $q \in \O_F$ is given and $x,y,z,t \in \O_F$ are the unknowns). We show that, in all likelihood, solving this equation requires the enumeration of a \textit{too large} set to be feasible, thereby making irrelevant a straightforward adaptation of the approach in~\cite{modLIPtotallyreal}.

Regardless of the choice of parameters, knowledge of a single signed message, i.e., a pair message/signature, produced by Kahrobaei-Koupparis digital signature scheme is sufficient to forge a valid signature for any other message.

Tweakable enciphering modes (TEMs) provide security in various storage and space-critical applications, including disk and file-based encryption and packet-based communication protocols. XCB-AES (originally introduced as XCBv2) is specified in the IEEE 1619.2 standard for encryption of sector-oriented storage media and comes with a formal security proof for block-aligned messages. In this work, we present the first plaintext recovery attack on XCB-AES $-$ the shared difference attack, demonstrating that the security of XCB-AES is fundamentally flawed. Our plaintext recovery attack is highly efficient and requires only two queries (one enciphering and one deciphering), breaking the claimed $\mathsf{vil\text{-}stprp}$, $\mathsf{stprp}$ as well as the basic $\mathsf{sprp}$ security. Our shared difference attack exploits an inherent property of polynomial hash functions called separability. We pinpoint the exact flaw in the security proof of XCB-AES, which arises from the separability of polynomial hash functions. We show that this vulnerability in the XCB design strategy has gone unnoticed for over 20 years and has been inadvertently replicated in many XCB-style TEM designs, including the IEEE 1619.2 standard XCB-AES. We also apply the shared difference attack to other TEMs based on XCB $-$ XCBv1, HCI, and MXCB, invalidating all of their security claims, and discuss some immediate countermeasures. Our findings are the first to highlight the need to reassess the present IEEE 1619.2 standard as well as the security and potential deployments of XCB-style TEMs.

We show that the key agreement scheme [IEEE TNSE, 1454--1468, 2004] is insecure against ephemeral secret leakage attack and smart card loss attack, not as claimed. This failure results from its simple encryption, in which only bitwise XOR operations are used to encrypt the messages.

The dual attacks on the Learning With Errors (LWE) problem are currently a subject of controversy. In particular, the results of [Matzov,2022], which claim to significantly lower the security level of CRYSTALS-Kyber, a lattice-based cryptosystem currently being standardized by NIST, are not widely accepted. The analysis behind their attack depends on a series of assumptions that, in certain scenarios, have been shown to contradict established theorems or well-tested heuristics [Ducas,Pulles,CRYPTO2023]. In this paper, we introduce a new dual lattice attack on LWE, drawing from ideas in coding theory. Our approach revisits the dual attack proposed by [Matzov,2022], replacing modulus switching with an efficient decoding algorithm. This decoding is achieved by generalizing polar codes over Zq, and we confirm their strong distortion properties through benchmarks. This modification enables a reduction from small-LWE to plain-LWE, with a notable decrease in the secret dimension. Additionally, we replace the enumeration step in the attack by assuming the secret is zero for the portion being enumerated, iterating this assumption over various choices for the enumeration part. We make an analysis of our attack without using the flawed independence assumptions used in [Matzov,2022] and we fully back up our analysis with experimental evidences. Lastly, we assess the complexity of our attack on CRYSTALS-Kyber; showing that the security levels for Kyber-512/768/1024 are 3.5/11.9/12.3 bits below the NIST requirements (143/207/272 bits) in the same nearest-neighbor cost model as in the Kyber submission. All in all the cost of our attack matches and even slightly beat in some cases the complexities originally claimed by the attack of [Matzov,2022].

Automated cryptanalysis has taken center stage in the arena of cryptanalysis since the pioneering work by Mouha et al., which showcased the power of Mixed Integer Linear Programming (MILP) in solving cryptanalysis problems that otherwise required significant effort. Since the inception, research in this area has moved in primarily two directions. One is to model more and more classical cryptanalysis tools as optimization problems to leverage the ease provided by state-of-the-art solvers. The other direction is to improve existing models to make them more efficient and/or accurate. The current work is an attempt to contribute to the latter. In this work, a general model referred to as DEEPAND has been devised to capture the correlation between AND gates in NLFSR-based lightweight block ciphers. DEEPAND builds upon and generalizes the idea of joint propagation of differences through AND gates captured using refined MILP modeling of TinyJAMBU by Saha et al. in FSE 2020. The proposed model has been applied to TinyJAMBU, KATAN, and KTANTAN and can detect correlations that were missed by earlier models. This leads to more accurate differential bounds for both the ciphers. In particular, a 384-round (full round as per earlier specification) Type-IV trail is found for TinyJAMBU with 14 active AND gates using the new model, while the refined model reported this figure to be 19. This also reaffirms the decision of the designers to increase the number of rounds from 384 to 640. Moreover, the model succeeds in searching a full-round Type-IV trail of TinyJAMBU keyed permutation P_1024 with probability 2^-105 (much greater than 2^-128). This reveals the non-random properties of P_1024, thereby showing it to be non-ideal. Hence, it cannot be expected to provide the same security levels as robust block ciphers. Further, the provable security of the TinyJAMBU AEAD scheme should be carefully revisited. Similarly, for the variants of KATAN, several previously reported trails are improved upon by employing the DEEPAND model. Moreover, in the related-key setting, the DEEPAND model is able to make a better 140-round boomerang distinguisher (for both the data and time complexity) in comparison to the previous boomerang attack by Isobe et al. in ACISP 2013. Furthermore, for enhanced applicability, we employ the DEEPAND model on another multiple-AND-based cipher, KTANTAN, in the related-key setting. Our analysis reveals practical differential distinguishers with low data and time complexities for all full-round KTANTAN variants. In summary, DEEPAND seems to capture the underlying correlation better when multiple AND gates are at play and can be adapted to other classes of ciphers as well.

The Internet of Things (IoT) has become integral to modern life, enabling smart cities, healthcare, and industrial automation. However, the increasing connectivity of IoT devices exposes them to various cyber threats, necessitating robust encryption methods. The PRESENT cipher, a lightweight block cipher, is well-suited for resource-constrained IoT environments, offering strong security with minimal computational overhead. This paper explores the application of deep learning (DL) techniques in cryptanalysis, specifically using an Aggregated Perceptron Group (APG) Model, which employs a Multi-Layer Perceptron (MLP) to predict input-output relations for each round of the PRESENT cipher’s encryption, excluding the key. This approach focuses solely on emulating the cipher's Substitution Permutation Network (SPN), capturing its essential structure and revealing the structural flaws in the way data is transformed through rounds. The models are chained together to generate the final ciphertext for any 64-bit plaintext with high accuracy, reducing the probability form a random guess of $2^{64}$. The results demonstrate the potential of DL models in cryptanalysis, providing insights into the security of lightweight ciphers in IoT communications and highlighting the practicality of deep learning for cryptographic analysis on standard computing systems.

The matrix code equivalence problem consists, given two matrix spaces $\mathcal{C},\mathcal{D} \subset \mathbb{F}_q^{m\times n}$ of dimension $k$, in finding invertible matrices $P\in\mathrm{GL}_m(\mathbb{F}_q)$ and $Q\in\mathrm{GL}_n(\mathbb{F}_q)$ such that $\mathcal{D}=P\mathcal{C} Q^{-1}$. Recent signature schemes such as MEDS and ALTEQ relate their security to the hardness of this problem. Recent works by Narayanan, Qiao and Tang on the one hand and by Ran and Samardjiska on the other hand tackle this problem. The former is restricted to the ``cubic'' case $k = m =n$ and succeeds in $\widetilde{\mathcal{O}}(q^{\frac k 2})$ operations. The latter is an algebraic attack on the general problem whose complexity is not fully understood and which succeeds only on $\mathcal{O}(1/q)$ instances. We present a novel algorithm which solves the problem in the general case. Our approach consists in reducing the problem to the matrix code conjugacy problem, \emph{i.e.} the case $P=Q$. For the latter problem, similarly to the permutation code equivalence problem in Hamming metric, a natural invariant based on the \emph{Hull} of the code can be used. Next, the equivalence of codes can be deduced using a usual list collision argument. For $k=m=n$, our algorithm achieves the same time complexity as Narayanan \emph{et al.} but with a lower space complexity. Moreover, ours extends to a much broader range of parameters.

The Learning With Errors (LWE) problem, introduced by Regev (STOC'05), is one of the fundamental problems in lattice-based cryptography, believed to be hard even for quantum adversaries. Regev (FOCS'02) showed that LWE reduces to the quantum Dihedral Coset Problem (DCP). Later, Brakerski, Kirshanova, Stehlé and Wen (PKC'18) showed that LWE reduces to a generalization known as the Extrapolated Dihedral Coset Problem (EDCP). We present a quasi-polynomial time quantum algorithm for the EDCP problems over power-of-two moduli using a quasi-polynomial number of samples, which also applies to the SLWE problem defined by Chen, Liu, and Zhandry (Eurocrypt'22). Our EDCP algorithm can be viewed as a provable variant to the "Simon-meets-Kuperberg" algorithm introduced by Bonnetain and Naya-Plasencia (Asiacrypt'18), adapted to the EDCP setting. We stress that our algorithm does not affect the security of LWE with standard parameters, as the reduction from standard LWE to EDCP limits the number of samples to be polynomial.

NIST standardized ML-KEM and ML-DSA as post-quantum key exchanges and digital signatures. Both schemes have already seen analysis with respect to side-channels, and first fully masked implementations of ML-DSA have been published. Previous attacks focused on unprotected implementations or assumed only hiding countermeasures to be in-place. Thus, in contrast to ML-KEM, the threat of side-channel attacks for protected ML-DSA implementations is mostly unclear. In this work, we analyze the side-channel vulnerability of masked ML-DSA implementations. We first systematically assess the vulnerability of several potential points of attacks in different leakage models using information theory. Then, we explain how an adversary could launch first, second, and higher-order attacks using a recently presented framework for side-channel information in lattice-based schemes. In this context, we propose a filtering technique that allows the framework to solve for the secret key from a large number of hints; this had previously been prevented by numerical instabilities. We simulate the presented attacks and discuss the relation to the information-theoretic analysis. Finally, we carry out relevant attacks on physical devices, discuss recent masked implementations, and instantiate a countermeasure against the most threatening attacks. The countermeasure mitigates the attacks with the highest noise-tolerance while having very little overhead. The results on the physical devices validate our simulations.

XCB, a tweakable enciphering mode, is part of IEEE Std. 1619.2 for shared storage media. We show that all versions of XCB are not secure through three plaintext recovery attacks. A key observation is that XCB behaves like an LRW1-type tweakable block cipher for single-block messages, which lacks CCA security. The first attack targets one-block XCB, using three queries to recover the plaintext. The second one requires four queries to recover the plaintext that excludes one block. The last one requires seven queries to recover the \emph{full} plaintext. The first attack applies to any scheme that follows the XCB structure, whereas the latter two attacks work on all versions of XCB, exploiting the separable property of the underlying universal hash function. To address these flaws, we propose the XCB* structure, an improved version of XCB that adds only two XOR operations. We prove that XCB* is STPRP-secure when using AXU hash functions, SPRPs, and a secure random-IV-based stream cipher.

The Rowhammer attack is a fault-injection technique leveraging the density of RAM modules to trigger persistent hardware bit flips that can be used for probing or modifying protected data. In this paper, we show that Falcon, the hash-and-sign signature scheme over NTRU lattices selected by NIST for standardization, is vulnerable to an attack using Rowhammer. Falcon's Gaussian sampler is the core component of its security, as it allows to provably decorrelate the short basis used for signing and the generated signatures. Other schemes, lacking this guarantee (such as NTRUSign, GGH or more recently Peregrine) were proven insecure. However, performing efficient and secure lattice Gaussian sampling has proved to be a difficult task, fraught with numerous potential vulnerabilities to be exploited. To avoid timing attacks, a common technique is to use distribution tables that are traversed to output a sample. The official Falcon implementation uses this technique, employing a hardcoded reverse cumulative distribution table (RCDT). Using Rowhammer, we target Falcon's RCDT to trigger a very small number of targeted bit flips, and prove that the resulting distribution is sufficiently skewed to perform a key recovery attack. Namely, we show that a single targeted bit flip suffices to fully recover the signing key, given a few hundred million signatures, with more bit flips enabling key recovery with fewer signatures. Interestingly, the Nguyen–Regev parallelepiped learning attack that broke NTRUSign, GGH and Peregrine does not readily adapt to this setting unless the number of bit flips is very large. However, we show that combining it with principal component analysis (PCA) yields a practical attack. This vulnerability can also be triggered with other types of persistent fault attacks on memory like optical faults. We suggest cheap countermeasures that largely mitigate it, including rejecting signatures that are unusually short.

We present a dataset of side-channel power measurements captured during pair-pointwise multiplication in the decapsulation procedure of the Kyber Key Encapsulation Mechanism (KEM). The dataset targets the pair-pointwise multiplication step in the NTT domain, a key computational component of Kyber. The dataset is collected using the reference implementation from the PQClean project. We hope the dataset helps in research in ``classical'' power analysis and deep learning-based side-channel attacks on post-quantum cryptography (PQC).

Ascon, a family of algorithms that support hashing and Authenticated Encryption with Associated Data (AEAD), is the final winner of the NIST Lightweight Cryptography Project. As a research hotspot, Ascon has received substantial third-party security evaluation. Among all the results of Ascon-128 (the primary recommendation of AEAD), the key recovery attack can only be achieved by reducing the initialization phase to 7 rounds or fewer, regardless of whether it violates the security claims made by the designers (i.e., misuse of the nonce or exceeding data limits $2^{64}$). In this paper, we, from two aspects (misuse-free setting and misused setting), improve the key recovery attack on Ascon-128 using the cube attack method. In one part, we present a faster method to recover the superpolies for a 64-dimensional cube in the output bits of the 7-round initialization, enabling us to recover the secret key with a time complexity of $2^{95.96}$ and a data complexity of $2^{64}$. Our 7-round key recovery attack, based on the full key space, greatly improves the time complexity, making it the best result to date. Additionally, we utilize several techniques to extend state recovery to key recovery, answering the open problem of transitioning from full state recovery in the encryption phase to key recovery for Ascon-128 (ToSc Vol 4, 2022). By combining encryption phase state recovery with initialization phase key recovery, we can achieve 8-round and 9-round initialization phase key recovery in the nonce misuse scenario, with time complexities of $2^{101}$ and $2^{123.92}$, respectively. This represents an improvement of two rounds over previous results in the misused setting. Our first key recovery attack is also applicable to Ascon-128a, achieving the same result. In cases where the full state, prior to the encryption phase, can be recovered in other Ascon AEAD modes, our second key recovery attack will also be useful. It is worth noting that this work does not threaten the security of the full 12 rounds Ascon, but we expect that our results provide new insights into the security of Ascon.

As cloud-based quantum computing services, such as those offered by D-Wave, become more popular for practical applications, privacy-preserving methods (such as obfuscation) are essential to address data security, privacy, and legal compliance concerns. Several efficient obfuscation methods have been proposed, which do not increase the time complexity of solving the obfuscated problem, for quantum optimization problems. These include {\em sign reversing}, {\em variable permutation}, and the combination of both methods assumed to provide greater protection. Unfortunately, sign reversing has already been shown to be insecure. We present two attacks on variable permutation and the combined method, where it is possible to efficiently recover the deobfuscated problem, particularly when given access to the obfuscated problem and its obfuscated solution, as a cloud-based quantum provider would have. Our attacks are in the context of an optimization problem of cryptanalysis of the Trivium cipher family, but our approach generalizes to other similarly structured problems. Our attacks are efficient and practical. Deobfuscating an optimization problem with \( n \) variables obfuscated with the combined method has a complexity of $O(n^2)$ compared to the complexity of $O\left( n \cdot n! \cdot 2^n \right)$ of the brute force attack. We provide an implementation of our attack; using a commodity laptop, our attack using the full Trivium cipher takes less than two minutes if optimized. We also present possible countermeasures to mitigate our attacks and bring attention to the need for further development in this area.

We show that the Negi-Kumar certificateless ring signature scheme [Wirel. Pers. Commun. 134(4): 1987-2011 (2024)] is insecure against forgery attack. The signer's public key $PK_j$ and secret key $PSK_j$ are simply invoked to compute the hash value $H_{2_j}=h_5(m_j\|PSK_j\|PK_j\|t_j)$, which cannot be retrieved by the verifier for checking their dependency. The explicit dependency between the public key and secret key is not properly used to construct some intractable problems, such as Elliptic Curve Discrete Logarithm (ECDL), Computational Diffie-Hellman (CDH), and Decisional Diffie-Hellman (DDH). An adversary can find an efficient signing algorithm functionally equivalent to the valid signing algorithm. The findings in this note could be helpful for the newcomers who are not familiar with the designing techniques for certificateless ring signature.

RIPEMD-128 is an ISO/IEC standard hash function based on a double-branch Merkle-Damgård structure. Its compression function includes two branches with distinct Boolean functions and message expansion permutations. To perform a collision attack, differential characteristics must be constructed simultaneously for both branches under the same message word difference, and the message modification order must align with conditions in both branches. These factors make collision attacks on (reduced) RIPEMD-128 highly challenging. In 2014, an attack on 40 steps of RIPEMD-128 was achieved by Wang with no state differences in round 3. In this work, we analyze message permutation properties and propose two new structures for creating message differences. These structures enable high-probability local collisions in both branches of round 3, extending the attack to more steps. Notably, the second structure can eliminate all state differences in round 3, allowing the attack to cover more than three whole rounds. To ensure practical attacks, we limit the number of conditions based on our message modification strategy and use multi-step message modification techniques to control more conditions. As a result, we successfully generate colliding message pairs for 46-step and 54-step reduced RIPEMD-128, with time complexities of approximately $2^{42}$ and $2^{54}$, respectively.

This paper studies quantum linear key-recovery attacks on block ciphers. The first such attacks were last-rounds attacks proposed by Kaplan et al. (ToSC 2016), which combine a linear distinguisher with a guess of a partial key. However, the most efficient classical attacks use the framework proposed by Collard et al. (ICISC 2007), which computes experimental correlations using the Fast Walsh-Hadamard Transform. Recently, Schrottenloher (CRYPTO 2023) proposed a quantum version of this technique, in which one uses the available data to create a quantum \emph{correlation state}, which is a superposition of subkey candidates where the amplitudes are the corresponding correlations. A limitation is that the good subkey is not marked in this state, and cannot be found easily. In this paper, we combine the correlation state with another distinguisher. From here, we can use Amplitude Amplification to recover the right key. We apply this idea to Feistel ciphers and exemplify different attack strategies on LOKI91 before applying our idea on the CAST-128 and CAST-256 ciphers. We demonstrate the approach with two kinds of distinguishers, quantum distinguishers based on Simon's algorithm and linear distinguishers. The resulting attacks outperform the previous Grover-meet-Simon attacks.

At the current state of the art, algebraic attacks are the most efficient method for finding preimages and collisions for arithmetization-oriented hash functions, such as the closely related primitives Poseidon/Poseidon2 and Neptune. In this paper, we revisit Gröbner basis (GB) attacks that exploit subspace trails to linearize some partial rounds, considering both sponge and compression modes. Starting from Poseidon's original security evaluation, we identified some inaccuracies in the model description that may lead to misestimated round requirements. Consequently, we reevaluate and improve the proposed attack strategy. We find that depending on the concrete instantiation, the original security analysis of Poseidon under- or overestimates the number of rounds needed for security. Moreover, we demonstrate that GB attacks leveraging subspace trails can outperform basic GB attacks for Poseidon/Poseidon2 and Neptune. We propose a variant of the previous attack strategy that exploits a crucial difference between Poseidon/Poseidon2 and Neptune: while Poseidon's inverse round functions have a high degree, Neptune's inverse external rounds maintain the same degree as the forward rounds. Using this new model, we demonstrate that Neptune's security in compression mode cannot be reduced to its security against the Constrained-Input-Constrained-Output (CICO) problem. To the best of our knowledge, this is the first time a concrete example has been provided where finding preimages is easier than solving the corresponding CICO problem. Our results emphasize the importance of considering the mode of operation in security analysis while confirming the overall security of Poseidon/Poseidon2 and Neptune against the presented algebraic attacks.

Poseidon and Poseidon2 are cryptographic hash functions designed for efficient zero-knowledge proof protocols and have been widely adopted in Ethereum applications. To encourage security research, the Ethereum Foundation announced a bounty program in November 2024 for breaking the Poseidon challenges, i.e. solving the CICO (Constrained Input, Constrained Output) problems for round-reduced Poseidon constructions. In this paper, we explain how to apply the Graeffe transform to univariate polynomial solving, enabling efficient interpolation attacks against Poseidon. We will provide an open-source code and details our approach for solving several challenges valued at $20000 in total. Compared to existing attacks, we improves 2^{13} and 2^{4.5} times in wall time and memory usage, respectively. For all challenges we solved, the cost of memory access turns out to be an essential barrier, which makes the security margin much larger than expected. We actually prove that the memory access cost for FFT grows as the 4/3-power of the input size, up to a logarithmic factor. This indicates the commonly used pseudo linear estimate may be overly conservative. This is very different from multivariate equation solving whose main bottleneck is linear algebra over finite fields. Thus, it might be preferable to choose parameters such that the best known attack is interpolation, as it presents more inherent hardness.

We show that the certificateless signature scheme [Veh. Commun. 47: 100763 (2024)] is insecure, because an adversary can launch forgery attack for any message. The signer's certificateless public key is not tightly bound to the system public key. The inherent flaw results in that the adversary can find an efficient signing algorithm functionally equivalent to the valid signing algorithm. The findings in this note could be helpful for newcomers who are not familiar with the designing techniques for certificateless signatures.

This paper explores the algebraic structure of the Poseidon and Poseidon2 permutations over NTT-friendly finite fields, with a focus on preimage recovery via root-finding techniques. We introduce an algorithm for efficiently identifying single roots of high-degree univariate polynomials that emerge from these constructions, based on the Graeffe transform and the tangent Graeffe method. Our approach is evaluated on reduced-round bounty instances of these permutations at various security levels, as proposed by the Ethereum Foundation, demonstrating practical effectiveness. These results yield new insights into the security of permutation-based cryptographic primitives instantiated over NTT-friendly prime fields.

Multivariate public key cryptography (MPKC) is one of the most promising alternatives to build quantum-resistant signature schemes, as evidenced in NIST's call for additional post-quantum signature schemes. The main assumption in MPKC is the hardness of the Multivariate Quadratic (MQ) problem, which seeks for a common root to a system of quadratic polynomials over a finite field. Although the Crossbred algorithm is among the most efficient algorithm to solve MQ over small fields, its complexity analysis stands on shaky ground. In particular, it is not clear for what parameters it works and under what assumptions. In this work, we provide a rigorous analysis of the Crossbred algorithm over any finite field. We provide a complete explanation of the series of admissible parameters proposed in previous literature and explicitly state the regularity assumptions required for its validity. Moreover, we show that the series does not tell the whole story, hence we propose an additional condition for Crossbred to work. Additionally, we define and characterize a notion of regularity for systems over a small field, which is one of the main building blocks in the series of admissible parameters.

Side-channel analysis (SCA) based on machine learning, particularly neural networks, has gained considerable attention in recent years. However, previous works predominantly focus on establishing connections between labels and related profiled traces. These approaches primarily capture label-related features and often overlook the connections between traces of the same label, resulting in the loss of some valuable information. Besides, the attack traces also contain valuable information that can be used in the training process to assist model learning. In this paper, we propose a profiled SCA approach based on contrastive learning named CL-SCA to address these issues. This approach extracts features by emphasizing the similarities among traces, thereby improving the effectiveness of key recovery while maintaining the advantages of the original SCA approach. Through experiments of different datasets from different platforms, we demonstrate that CL-SCA significantly outperforms other approaches. Moreover, by incorporating attack traces into the training process using our approach, we can further enhance its performance. This extension can improve the effectiveness of key recovery, which is fully verified through experiments on different datasets.

The search rank-2 module Lattice Isomorphism Problem (smLIP), over a cyclotomic ring of degree a power of two, can be reduced to an instance of the Lattice Isomorphism Problem (LIP) of at most half the rank if an adversary knows a nontrivial automorphism of the underlying integer lattice. Knowledge of such a nontrivial automorphism speeds up the key recovery attack on HAWK at least quadratically, which would halve the number of security bits. Luo et al. (ASIACRYPT 2024) recently found an automorphism that breaks omSVP, the initial underlying hardness assumption of HAWK. The team of HAWK amended the definition of omSVP to include this so-called symplectic automorphism in their submission to the second round of NIST's standardization of additional signatures. This work provides confidence in the soundness of this updated definition, assuming smLIP is hard, since there are plausibly no more trivial automorphisms that allow winning the omSVP game easily. Although this work does not affect the security of HAWK, it opens up a new attack avenue involving the automorphism group that may be theoretically interesting on its own.

Side-channel analysis (SCA) has become a crucial area in ensuring the security of hardware implementations against potential vulnerabilities. With advancements in machine learning (ML) and artificial intelligence (AI), neural network(NN)-assisted techniques for SCA have demonstrated significant effectiveness. However, a fundamental question remains unanswered: are traces corresponding to different subkeys equally hard to attack? This paper addresses this issue by leveraging explainable AI techniques to analyze the hardness levels and, consequently, the root cause of hardness. To systematically investigate this problem, we reintroduce hardness metrics in SCA, which have been known to the ML community. Those metrics include query hardness (QH), log odds (LO), and matrix-based Rényi entropy (MRE). The challenge in this regard is that (virtually all) hardness metrics in ML cannot be adopted as they are. This is because ML and SCA metrics have conflicting goals, namely boosting accuracy and rank. Through careful experimentation, we identify the shortcomings of QH and LO in SCA and recommend MRE as a suitable hardness metric for SCA. We also study how hardness has been seen in SCA, where recent work has suggested the influence of class “labels” on the attack difficulty. Employing rigorous evaluation, our paper demonstrates that no statistically significant evidence can be found to support this claim. This leads us to the question of how much traces’ time samples affect the inherent hardness in distinguishing key candidates. Our novel explainable AI (XAI) approach not only answers this, but also makes a link between hardness and rank as the common performance metric in SCA. Our findings indicate that hardness values are influenced mainly by time samples rather than specific key labels. Furthermore, we examine whether hardness captures intrinsic properties of the traces, specifically, their lack of separability in feature space due to their inherent similarities. To this end, we introduce, for the first time in the context of SCA, the use of maximum mean discrepancy (MMD) as a principled metric. MMD effectively links trace hardness with attack difficulty by quantifying distributional differences induced by traces’ time samples. In addition to visualization techniques showing the root cause of hardness based on MMD, we employ XAI to explain the connection between MMD and attack hardness. Our proposed methodology enhances the understanding of attack difficulty in SCA and contributes to developing more robust evaluation metrics for profiling attacks.

Spiking neural networks gain attention due to low power properties and event-based operation, making them suitable for usage in resource constrained embedded devices. Such edge devices allow physical access opening the door for side-channel analysis. In this work, we reverse engineer the parameters of a feed-forward spiking neural network implementation with correlation power analysis. Localized measurements of electro-magnetic emanations enable our attack, despite inherent parallelism and the resulting algorithmic noise of the network. We provide a methodology to extract valuable parameters of integrate-and-fire neurons in all layers, as well as the layer sizes.

Rivest, Shamir, and Adleman published the RSA cryptosystem in 1978, which has been widely used over the last four decades. The security of RSA is based on the difficulty of factoring large integers $N = pq$, where $p$ and $q$ are prime numbers. The public exponent $e$ and the private exponent $d$ are related by the equation $ed - k(p-1)(q-1) = 1$. Recently, Cotan and Teseleanu (NordSec 2023) introduced a variant of RSA, where the public exponent $e$ and the private exponent $d$ satisfy the equation $ed - k(p^n-1)(q^n-1) = 1$ for some positive integer $n$. In this paper, we study the general equation $eu - (p^n - 1)(q^n - 1)v = w$ with positive integers $u$ and $v$, and $w\in \mathbb{Z}$. We show that, given the public parameters $N$ and $e$, one can recover $u$ and $v$ and factor the modulus $N$ in polynomial time by combining continued fractions with Coppersmith's algorithm which relies on lattice reduction techniques, under specific conditions on $u$, $v$, and $w$. Furthermore, we show that if the private exponent $d$ in an RSA-like cryptosystem is either small or too large, then $N$ can be factored in polynomial time. This attack applies to the standard RSA cryptosystem.

In 2024, based on the cubic Pell curve, Nitaj and Seck proposed a variant of the RSA cryptosystem where the modulus is in the form $N=p^rq^s$, and the public key $e$ and private key $d$ satisfy the equation $ed\equiv 1\pmod{(p-1)^2(q-1)^2}$. They showed that $N$ can be factored when $d$ is less than a certain bound that depends on $r$ and $s$ in the situation $rs\geq 2$, which is not extendable to $r=s=1$. In this paper, we propose a cryptanalysis of this scheme in the situation $r=s=1$, and give an explicit bound for $d$ that makes the scheme insecure. The method is based on Coppersmith's method and lattice reduction.

Mitaka is a variant of Falcon, which is one of the three postquantum signature schemes selected by the NIST for standardization. Introduced in 2021, Mitaka offers a simpler, parallelizable, and maskable version of Falcon. Our article focuses on its physical security, and our results are threefold. Firstly, we enhance a known profiled side-channel attack on an unprotected Mitaka implementation by a factor of 512. Secondly, we analyze the masked implementation of Mitaka described in the reference article, which incorporates a different sampler and a protective gadget. We expose the first side-channel flaw on this sampler. This flaw enables to break the masked implementation with a first-order side-channel attack. In this scenario, the key can be recovered using only three times more traces compared to the attack on the unprotected implementation. Finally, we discuss and recommend new countermeasures to mitigate these attacks.

At Africacrypt 2022, Gupta et al. introduced FUTURE, a 64-bit lightweight block cipher based on an MDS matrix and designed in an SPN structure, with a focus on achieving single-cycle encryption and low implementation cost, especially in unrolled architectures. While the designers evaluated its security under various attack models, they did not consider related-key cryptanalysis. In this work, we address this gap by analyzing the security of FUTURE in the related-key setting using techniques based on Mixed Integer Linear Programming (MILP). We first propose a simplified and generalizable approach for applying MILP to model any MDS or near-MDS-based cipher that follows the substitution-permutation paradigm. Using our MILP framework, we construct an 8-round related-key distinguisher on FUTURE, requiring $2^{58}$ plaintexts, $2^{58}$ \xor operations, and negligible memory. We further identify a full-round (i.e., 10 rounds) boomerang distinguisher with a probability of $2^{-45.8}$, enabling a distinguishing attack with $2^{48}$ data and time complexity. In addition, we develop a full-round key recovery attack on FUTURE with data, time, and memory complexities of $2^{18}$, $2^{70}$, and $2^{64}$, respectively. Although all known single-key attacks remain impractical (with time complexities of at least $2^{112}$), our results demonstrate a full-round cryptanalysis of FUTURE in the related-key setting, thereby challenging its claimed security guarantees.

Homomorphic encryption provides many opportunities for privacy-aware processing, including with methods related to machine learning. Many of our existing cryptographic methods have been shown in the past to be susceptible to side channel attacks. With these, the implementation of the cryptographic methods can reveal information about the private keys used, the result, or even the original plaintext. An example of this includes the processing of the RSA exponent using the Montgomery method, and where 0's and 1's differ in their processing time for modular exponentiation. With FHE, we typically use lattice methods, and which can have particular problems in their implementation in relation to side channel leakage. This paper aims to outline a range of weaknesses within FHE implementations as related to side channel analysis. It outlines a categorization for side-channel analysis, some case studies, and mitigation strategies.

We explore the application of the QUBO and Ising models to the integer factorization problem with implications for the security of public-key algorithms such as RSA. A key contribution is a program that applies existing algorithms to parameterize and simulate integer factorization through an Ising model in order to replicate previous works. Due to limited access to quantum hardware, we use classical heuristic methods to approximate solutions.

The growing adoption of secure multi-party computation (MPC) has driven the development of efficient symmetric key primitives tailored for MPC. Recent advances, such as the alternating moduli paradigm, have shown promise but leave room for cryptographic and practical improvements. In this paper, we analyze a family of weak pseudorandom functions (wPRF) proposed at Crypto 2024, focusing on their One-to-One parameter sets. We demonstrate that these configurations fail to achieve their intended one-to-one mappings and exploit this observation to develop an efficient key recovery attack. Our analysis reveals critical vulnerabilities, reducing the complexity of key recovery to $\mathcal{O}(2^{\lambda/2} \log_2 \lambda)$ for the Standard One-to-One wPRF and $\mathcal{O}(2^{0.84\lambda})$ for the Reversed Moduli variant -- both substantially below their claimed $\lambda$-bit security. We validate our findings through experimental evaluation, confirming alignment between predicted and observed attack complexities.

Side-channel analysis remains a critical threat to public-key cryptographic implementations. Simple Power Analysis (SPA) techniques can extract secret keys from a single power trace, often using clustering-based classification methods. However, traces captured in real-world environments often suffer from misalignment and variable trace lengths due to unstable clocks and random delays. As a result, clustering methods are required to use alignment methods that may alter the original information of the traces. To address this problem, this work proposes Dynamic Time Classification (DTC) as an alternative approach to classify cryptographic operations in SPA based on Dynamic Time Warping. Unlike clustering methods, DTC inherently compares power traces without requiring fixed-length segments, which greatly improved the adaptability to unequal traces and thus allows us to classify different operations relatively stably. Experimental results on public-key cryptographic algorithms and post-quantum algorithm implementations show that DTC are no less accurate than clustering methods and are more robust to timing variations. This work also systematically divides the features of different operations and explores the effects of different SPA methods on different types of feature. This work also conducts experiments with and without random delays for different categories, compares the experimental accuracy of different alignment methods, and discusses the feasibility of DTW as a preprocessing method.

Soft Analytical Side Channel Attacks (SASCA) are a powerful family of Side Channel Attacks (SCA) that allows the recovery of secret values with only a small number of traces. Their effectiveness lies in the Belief Propagation (BP) algorithm, which enables efficient computation of the marginal distributions of intermediate values. Post-quantum schemes such as Kyber, and more recently, Hamming Quasi-Cyclic (HQC), have been targets of SASCA. Previous SASCA on HQC focused on Reed-Solomon (RS) codes and successfully retrieved the shared key with a high success rate for high noise levels using a single trace. In this work, we present new SASCA on HQC, where both the shared key and the secret key are targeted. Our attacks are realized on simulations. Unlike the previous SASCA, we take a closer look at the Reed-Muller (RM) code. The advantage of this choice is that the RM decoder is applied before the RS decoder, enabling attacks targeting both the secret key and shared key. We build a factor graph of the Fast Hadamard Transform (FHT) function from the HQC reference implementation of April 2023. The information recovered from BP allows us to retrieve the shared key with a single trace. In addition to the previous SASCA targeting HQC, we also manage to recover the secret key with two different chosen ciphertext attacks. One of them requires a single trace and is successful until high noise levels.

This paper presents a side-channel attack targeting the LESS and CROSS post-quantum digital signature schemes, resulting in full key recovery for both. These schemes have advanced to the second round of NIST’s call for additional signatures. By leveraging correlation power analysis and horizontal attacks, we are able to recover the secret key by observing the power consumption during the multiplication of an ephemeral secret vector with a public matrix. The attack path is enabled by the presence of a direct link between the secret key elements and the ephemeral secret, given correct responses. This attack targets version 1.2 of both schemes. In both settings we can recover the secret key in a single trace for the NIST’s security level I parameter set. Additionally, we propose improvements to the existing horizontal attack on CROSS, reducing the required rounds that need to be observed by an order of magnitude for the same parameter sets.

We present an attack on the Abstract Datagram Network Layer (ADNL) protocol used in The Open Network (TON), currently the tenth largest blockchain by market capitalization. In its TCP variant, ADNL secures communication between clients and specialized nodes called liteservers, which provide access to blockchain data. We identify two cryptographic design flaws in this protocol: a handshake that permits session-key replay and a non-standard integrity mechanism whose security critically depends on message confidentiality. We transform these vulnerabilities into an efficient plaintext-recovery attack by exploiting two ADNL communication patterns, allowing message reordering across replayed sessions. We then develop a plaintext model for this scenario and construct an efficient algorithm that recovers the keystream using a fraction of known plaintexts and a handful of replays. We implement our attack and show that an attacker intercepting the communication between a TON liteserver and a widely deployed ADNL client can recover the keystream used to encrypt server responses by performing eight connection replays to the server. This allows the decryption of sensitive data, such as account balances and user activity patterns. Additionally, the attacker can modify server responses to manipulate blockchain information displayed to the client, including account balances and asset prices.

We show that the authentication method [Future Gener. Comput. Syst. 158: 516-529 (2024)] cannot be practically implemented, because the signature scheme is insecure against certificateless public key replacement forgery attack. The explicit dependency between the certificateless public key and secret key is not properly used to construct some intractable problems, such as Elliptic Curve Discrete Logarithm (ECDL). An adversary can find an efficient signing algorithm functionally equivalent to the valid signing algorithm. We also correct some typos in the original presentation.

The isogeny-based post-quantum digital signature algorithm SQIsign offers the most compact key and signature sizes among all candidates in the ongoing NIST call for additional post-quantum signature algorithms. To the best of our knowledge, we present the first Simple Power Analysis (SPA) side-channel attack on SQIsign, demonstrating its feasibility for key recovery. Our attack specifically targets secret-dependent computations within Cornacchia's algorithm, a fundamental component of SQIsign's quaternion module. At the core of this algorithm, a secret-derived yet ephemeral exponent is used in a modular exponentiation subroutine. By performing SPA on the modular exponentiation, we successfully recover this ephemeral exponent. We then develop a method to show how this leaked exponent can be exploited to ultimately reconstruct the secret signing key of SQIsign. Our findings emphasize the critical need for side-channel-resistant implementations of SQIsign, highlighting previously unexplored vulnerabilities in its design.

Arithmetization-Oriented (AO) symmetric primitives play an important role in the efficiency and security of zero-knowledge (ZK) proof systems. The design and cryptanalysis of AO symmetric-key primitives is a new topic particularly focusing on algebraic aspects. An efficient AO hash function aims at lowering the multiplicative complexity in the arithmetic circuit of the hash function over a suitable finite field. The AO hash function Anemoi was proposed in CRYPTO 2023. In this work we present an in-depth Groebner basis (GB) cryptanalysis of Anemoi over GF(p). The main aim of any GB cryptanalysis is to obtain a well-structured set of polynomials representing the target primitive, and finally solve this system of polynomials using an efficient algorithm. We propose a new polynomial modelling for Anemoi that we call ACICO. We show that using ACICO one can obtain a GB defined by a well-structured set of polynomials. Moreover, by utilising ACICO we can prove the exact complexity of the Groebner basis computation (w.r.t Buchberger's algorithm) in the cryptanalysis of Anemoi. The structured GB further allows us to prove the dimension of the quotient space which was conjectured in a recently published work. Afterwards, we provide the complexity analysis for computing the variety (or the solutions) of the GB polynomial system (corresponding to Anemoi) which is the final step in GB cryptanalysis, by using known approaches. In particular, we show that GB polynomial structure allows us to use the Wiedemann algorithm and improve the efficiency of cryptanalysis compared to previous works. Our GB cryptanalysis is applicable to more than two branches (a parameter in Anemoi), while the previously published results showed cryptanalysis only for two branches. Our complexity analysis implies that the security of Anemoi should not be relied upon the GB computation. We also address an important mathematical question in GB cryptanalysis of Anemoi namely, does the Anemoi polynomial system has a Shape form?, positively. By proving this we guarantee that upon application of basis conversion method like FGLM one can obtain a convenient system of polynomials that is easy to solve.

We show that the data classification scheme [IEEE Trans. Sustain. Comput., 2023, 8(4), 652-669)] failed to check the compatibility of encoding algorithm and homomorphic encryption algorithm. Some calculations should be revised to ensure all operands are first encoded using the same scaling factors. The canonical embedding map depending on the natural projection should be explicitly arranged so as to construct an efficient decoding algorithm.

Secure multi-party computation (MPC) enables parties to compute a function over private inputs while maintaining confidentiality. Although MPC has advanced significantly and attracts a growing industry interest, open-source implementations are still at an early stage, with no production-ready code and a poor understanding of their actual security guarantees. In this work, we study the real-world security of modern MPC implementations, focusing on the SPDZ protocol (Damgård et al., CRYPTO 2012, ESORICS 2013), which provides security against malicious adversaries when all-but-one of the participants may be corrupted. We identify a novel type of MAC key leakage in the MAC check protocol of SPDZ, which can be exploited in concurrent, multi-threaded settings, compromising output integrity and, in some cases, input privacy. In our analysis of three SPDZ implementations (MP-SPDZ, SCALE-MAMBA, and FRESCO), two are vulnerable to this attack, while we also uncover further issues and vulnerabilities with all implementations. We propose mitigation strategies and some recommendations for researchers, developers and users, which we hope can bring more awareness to these issues and avoid them reoccurring in future.

Indistinguishability is a fundamental principle of cryptographic security, crucial for securing data transmitted between Internet of Things (IoT) devices. This principle ensures that an attacker cannot distinguish between the encrypted data, also known as ciphertext, and random data or the ciphertexts of the two messages encrypted with the same key. This research investigates the ability of machine learning (ML) in assessing indistinguishability property in encryption systems, with a focus on lightweight ciphers. As our first case study, we consider the SPECK32/64 and SIMON32/64 lightweight block ciphers, designed for IoT devices operating under significant energy constraints. In this research, we introduce MIND-Crypt, a novel ML-based framework designed to assess the cryptographic indistinguishability of lightweight block ciphers, specifically the SPECK32/64 and SIMON32/64 encryption algorithm in CBC mode (Cipher Block Chaining), under Known Plaintext Attacks (KPA). Our approach involves training ML models using ciphertexts from two plaintext messages encrypted with same key to determine whether ML algorithms can identify meaningful cryptographic patterns or leakage. Our experiments show that modern ML techniques consistently achieve accuracy equivalent to random guessing, indicating that no statistically exploitable patterns exists in the ciphertexts generated by considered lightweight block ciphers. Furthermore, we demonstrate that in ML algorithms with all the possible combinations of the ciphertexts for given plaintext messages reflects memorization rather than generalization to unseen ciphertexts. Collectively, these findings suggest that existing block ciphers have secure cryptographic designs against ML-based indistinguishability assessments, reinforcing their security even under round-reduced conditions.

At CANS 2024, Zbudila et al. presented MaSTer, a maliciously secure multi-party computation protocol for truncation. It allows adversaries to manipulate outputs with a bounded additive error while avoiding detection with a certain probability. In this work, we analyse the broader implications of adversarial exploitation in probabilistic truncation protocols, specifically in relation to MaSTer. We propose three attack strategies aimed at inducing misclassification in deep neural network (DNN) inference. Our empirical evaluation across multiple datasets demonstrates that while adversarial influence remains negligible under realistic constraints, certain configurations and network architectures exhibit increased vulnerability. By improving the understanding of the risks associated with probabilistic truncation protocols in privacy-preserving machine learning, our work demonstrates that the MaSTer protocol is robust in realistic settings.

Differential Fault Attacks (DFAs) have recently emerged as a significant threat against stream ciphers specifically designed for Hybrid Homomorphic Encryption (HHE). In this work, we propose DFAs on the $\textsf{FRAST}$ cipher, which is a cipher specifically tailored for Torus-based Fully Homomorphic Encryption (TFHE). The round function of $\textsf{FRAST}$ employs random S-boxes to minimize the number of rounds, and can be efficiently evaluated in TFHE. With our specific key recovery strategy, we can mount the DFA with a few faults. Under the assumption of precise fault injection, our DFA can recover the key within one second using just 4 or 6 faults. When discarding the assumption and considering a more practical fault model, we can still achieve key recovery in a few minutes without increasing the number of faults. To the best of our knowledge, this is the first third-party cryptanalysis on $\textsf{FRAST}$. We also explored countermeasures to protect $\textsf{FRAST}$. Our analysis revealed that negacyclic S-boxes, a key component of TFHE-friendly ciphers, are unsuitable for incorporating linear structures to resist DFA. Consequently, we recommend removing the negacyclic restriction in the penultimate round of FRAST and introducing non-zero linear structures into the S-boxes of the last two rounds. We believe that our work will provide valuable insights for the design of TFHE-friendly ciphers.

We propose a new method for retrieving the algebraic structure of a generic alternant code given an arbitrary generator matrix, provided certain conditions are met. We then discuss how this challenges the security of the McEliece cryptosystem instantiated with this family of codes. The central object of our work is the quadratic hull related to a linear code, defined as the intersection of all quadrics passing through the columns of a given generator or parity-check matrix, where the columns are considered as points in the affine or projective space. The geometric properties of this object reveal important information about the internal algebraic structure of the code. This is particularly evident in the case of generalized Reed-Solomon codes, whose quadratic hull is deeply linked to a well-known algebraic variety called the rational normal curve. By utilizing the concept of Weil restriction of affine varieties, we demonstrate that the quadratic hull of a generic dual alternant code inherits many interesting features from the rational normal curve, on account of the fact that alternant codes are subfield-subcodes of generalized Reed-Solomon codes. If the rate of the generic alternant code is sufficiently high, this allows us to construct a polynomial-time algorithm for retrieving the underlying generalized Reed-Solomon code from which the alternant code is defined, which leads to an efficient key-recovery attack against the McEliece cryptosystem when instantiated with this class of codes. Finally, we discuss the generalization of this approach to Algebraic-Geometry codes and Goppa codes.

Side-channel analysis recovers a secret by exploiting the key-dependent characteristics of the leakages. Practical techniques, such as Distance-of-Means analysis (DoM), Kolmogorov-Smirnov analysis (KSA) and Cramér-von Mises analysis (CvMA), provide valuable insights about the secret from the indirect perspectives of statistical moment and cumulative distribution function (CDF) respectively, circumventing the direct and costly estimation of leakage probability densities and therefore enabling wider applicability in practice. Though both the perspectives are informative, their relationships in the context of side-channel analysis remain unclear. In other words, the fundamental questions of "which one is better?'' and ``why and under what circumstances?" leave as open problems. In this paper, we introduce the probability-probability (PP) plot in statistics as a common framework for explaining the mathematical foundations of CDF-based techniques, which facilitates an intuitive understanding of different variant strategies. Then, inspired by the growth pattern of the PP curve, we propose a novel distinguisher based on the famous Mann-Kendall test, where measurements are managed with ordinality and nominality. This goodness-of-fit test checks whether a key-dependent binary sequence originates from a random binomial distribution, by efficiently searching potential label clusters. Finally, we explore the symmetry and dual counterpart of CDF in mathematics, introducing the quantile-quantile (QQ) plot and develop an interesting technique based on the inverse cumulative distribution function (ICDF). We present a general discussion of its bridging role, regarding detail capture as well as signal-to-noise ratio (SNR). On this basis, we establish the relationships among moment-based, ICDF-based, and CDF-based techniques, which additionally allows for bounding the security level of the CDF-based techniques using well-established metrics that are originally proposed for evaluating the traditional moment-based family. Experiments across various settings validate our theoretical findings and demonstrate the effectiveness of the two proposed distinguishers.

We show that the data aggregation scheme [IEEE TDSC, 2023, 20(3), 2011-2024] is flawed, because the signer only signs a part of data, not the whole data. An adversary can replace the unsigned component to cheat the verifier. To frustrate this attack, all components of the target data should be concatenated together and then be hashed and signed, so as to ensure that the signature verification can prove the whole message integrity.

Multivariate public key cryptography (MPKC) is considered a promising candidate for post-quantum cryptography, with its security relying on the hardness of solving systems of multivariate quadratic equations. Among MPKC schemes, the unbalanced oil and vinegar (UOV) and its variants have been actively studied. Pébereau and Luyten showed that the Kipnis–Shamir attack and the singular point attack can be described within the same framework using the Jacobian matrix. In this study, we demonstrate that the rectangular MinRank attack can also be described within this framework. Furthermore, by leveraging this framework, we extend the feasible target ranks of the rectangular MinRank attack and use this extended attack to analyze the security of UOV and its variants. In conclusion, we confirm that the currently proposed parameters for UOV, MAYO, QR-UOV, and SNOVA are resistant to this attack.

Scloud$^{+}$ is a next-generation post-quantum key encapsulation mechanism that combines unstructured-LWE and a ternary key encoding technique to achieve efficient lattice cryptographic operations while eliminating traditional ring structure constraints. Despite its rigorously formalized theoretical security, its practical deployment faces side-channel threats, notably Correlation Power Analysis (CPA) attacks. This paper systematically investigates the physical security of its core ciphertext-key matrix multiplication module by proposing a CPA framework that integrates instruction-level timing analysis. A SoST (Sum of Squared T-values) model, inspired by multi-group Welch's t-test, is used to analyze the Hamming weight leakage during ciphertext loading. At the same time, dynamic sampling windows, combined with processor pipeline modeling, are employed to pinpoint critical leakage intervals. Exploiting the characteristics of ternary keys, an iterative recovery strategy is devised: following a predefined scan order, the candidate set $\{-1, 0, 1\}$ and partial intermediate sums are used to construct a Hamming weight model for hypothesized leakage vectors. Pearson correlation analysis and trace-count stabilization are applied within each dynamic sampling window to determine the optimal estimate for each key element. Experiments targeting 4800 key elements, illustrated through a detailed analysis of the first 32 elements, demonstrate high recovery accuracy with no more than 15 traces per element, indicating high efficiency and stability that can extend to the full key reconstruction. To thwart such CPA attacks, we have further designed and implemented a first‐order arithmetic masking countermeasure that splits the original ternary secret key into two subkeys, thereby expanding the attacker's key hypothesis space and significantly enhancing side‐channel resilience. Our results demonstrate that Scloud$^{+}$ remains vulnerable to side‐channel exploits at the implementation level, highlighting the urgent need to integrate appropriate protections into its standardization process.

Private Join and Compute (PJC) is a two-party protocol recently proposed by Google for various use-cases, including ad conversion (Asiacrypt 2021) and which generalizes their deployed private set intersection sum (PSI-SUM) protocol (EuroS&P 2020). PJC allows two parties, each holding a key-value database, to privately evaluate the inner product of the values whose keys lie in the intersection. While the functionality output is not typically considered in the security model of the MPC literature, it may pose real-world privacy risks, thus raising concerns about the potential deployment of protocols like PJC. In this work, we analyze the risks associated with the PJC functionality output. We consider an adversary that is a participating party of PJC and describe four practical attacks that break the other party's input privacy, and which are able to recover both membership of keys in the intersection and their associated values. Our attacks consider the privacy threats associated with deployment and highlight the need to include the functionality output as part of the MPC security model.

In this paper, we focus on SM4, a widely used and standardized Chinese block cipher. After revisiting the previously proposed optimal 19-round differential characteristic, we observe that its applicability in differential attacks is limited by a reduced pre-sieving probability, causing the time complexity to exceed that of brute force. To overcome this issue, we employ an automated search approach to identify more promising optimal 19-round differential characteristics. By translating key properties relevant to key recovery into Boolean expressions, we uncover three structural properties common to all optimal 19-round characteristics. While these properties dictate the overall probability of the resulting 19-round distinguishers, their varying pre-sieving probabilities influence their practical effectiveness in differential attacks. Using Boolean encodings, we identify four representative key-recovery-friendly differential characteristics. We then conduct an in-depth analysis of one such characteristic and demonstrate that, when evaluated under both the hypothesis testing paradigm and the key ranking paradigm, the proposed attack requires slightly more data than existing 23-round attacks. Nonetheless, it achieves lower time and memory complexities and ensures a higher success probability, offering a valuable new avenue for differential cryptanalysis of SM4. We believe our findings enhance the understanding of SM4's differential structure and provide a solid foundation for future research on advanced key-recovery techniques that leverage these newly identified structural properties and differential characteristics.

The security of the Elliptic Curve Digital Signature Algorithm (ECDSA) depends on the uniqueness and secrecy of the nonce, which is used in each signature. While it is well understood that nonce $k$ reuse across two distinct messages can leak the private key, we show that even if a distinct value is used for $k_2$, where an affine relationship exists in the form of: \(k_m = a \cdot k_n + b\), we can also recover the private key. Our method requires only two signatures (even over the same message) and relies purely on algebra, with no need for lattice reduction or brute-force search(if the relationship, or offset, is known). To our knowledge, this is the first closed-form derivation of the ECDSA private key from only two signatures over the same message, under a known affine relationship between nonces.

We show that here standard decoding algorithms for generic linear codes over a finite field can speeded up by a factor which is essentially the size of the finite field by reducing it to a low weight codeword problem and working in the relevant projective space. We apply this technique to SDitH and demonstrate that the parameters of the original submission fail to meet the security requirements set by NIST. However, the updated version, released after the discovery of this attack, is no longer challenged by our attack.

Deep learning-based side-channel analysis is an extremely powerful option for profiling side-channel attacks. However, to perform well, one needs to select the neural network model and training time hyperparameters carefully. While many works investigated these aspects, random search could still be considered the current state-of-the-art. Unfortunately, random search has drawbacks, since the chances of finding a good architecture significantly drop when considering more complex targets. In this paper, we propose a novel neural architecture search approach for SCA based on grammatical evolution - SCAGE. We define a custom SCA grammar that allows us to find well-performing and potentially unconventional architectures. We conduct experiments on four datasets, considering both synchronized and desynchronized versions, as well as using feature intervals or raw traces. Our results show SCAGE to perform extremely well in all settings, outperforming random search and related works in most of the considered scenarios.

At CRYPTO 2019, Gohr pioneered the integration of differential cryptanalysis with neural networks, demonstrating significant advantages over traditional distinguishers. Subsequently, at Inscrypt 2020, Su et al. proposed the concept of constructing polyhedral differential neural distinguishers by leveraging multiple effective input differences. More recently, at FSE 2024, Bellini et al. introduced a general-purpose tool for automating the training of single-key differential neural distinguishers for various block ciphers. Inspired by this body of work, we aim to extend automated search techniques to related-key differential neural distinguishers, enabling the discovery of effective input differences and key differences for such distinguishers. To achieve this, we employ a genetic optimization algorithm to identify effective differential combinations. To validate the efficacy of our method, we apply it to the Simeck and Simon cipher families, successfully identifying effective differential combinations for the three variants of Simeck and ten variants of Simon. Furthermore, inspired by the concept of polyhedral neural distinguishers, we adopt a novel data format that leverages multiple distinct input differences and key differences to construct positive and negative samples, providing the neural network with a richer set of features. Our approach not only identify high-quality distinguishers for previously unexplored cipher variants but also achieve higher accuracy for related-key differential neural distinguishers compared to the state-of-the-art.