Paper 2025/904

The Security of ML-DSA against Fault-Injection Attacks

Abstract

Deterministic signatures are often used to mitigate the risks associated with poor-quality randomness, where the randomness in the signing process is generated by a pseudorandom function that takes a message as input. However, some studies have shown that such signatures are vulnerable to fault-injection attacks. To strike a balance, recent signature schemes often adopt "hedged" randomness generation, where the pseudorandom function takes both a message and a nonce as input. Aranha et al. (EUROCRYPT 2020) investigated the security of hedged Fiat-Shamir signatures against 1-bit faults and demonstrated security for certain types of bit-tampering faults. Grilo et al. (ASIACRYPT 2021) extended this proof to the quantum random oracle model. Last year, NIST standardized the lattice-based signature scheme ML-DSA, which adopts the hedged Fiat-Shamir with aborts. However, existing security proofs against bit-tampering faults do not directly apply, as Aranha et al. left this as an open problem. To address this gap, we analyze the security of ML-DSA against multi-bit fault-injection attacks. We provide a formal proof of security for a specific class of intermediate values, showing that faults at these points cannot be exploited. Furthermore, to highlight the infeasibility of stronger fault resilience, we present key-recovery attacks that exploit signatures generated under fault injection at the other intermediate values.