Paper 2022/106 Original: eprint.iacr.org/2022/106 ↗

Profiling Side-Channel Attacks on Dilithium: A Small Bit-Fiddling Leak Breaks It All

Abstract

We present an end-to-end (equivalent) key recovery attack on the Dilithium lattice-based signature scheme, one of the top contenders in the NIST postquantum cryptography competition. The attack is based on a small side-channel leakage we identified in a bit unpacking procedure inside Dilithium signature generation. We then combine machine-learning based profiling with various algorithmic techniques, including least squares regression and integer linear programming, in order to leverage this small leakage into essentially full key recovery: we manage to recover, from a moderate number of side-channel traces, enough information to sign arbitrary messages. We confirm the practicality of our technique using concrete experiments against the ARM Cortext-M4 implementation of Dilithium, and verify that our attack is robust to real-world conditions such as noisy power measurements. This attack appears difficult to protect against reliably without strong side-channel countermeasures such as masking of the entire signing algorithm, and underscores the necessity of implementing such countermeasures despite their known high cost.