International Association
for Cryptologic Research

The Fiat-Shamir transform is one of the most widely applied methods for secure signature construction. Fiat-Shamir starts with an interactive zero-knowledge identification protocol and transforms this via a hash function into a non-interactive signature. The protocol's zero-knowledge property ensures that a signature does not leak information on its secret key $\vec s$, which is achieved by blinding $\vec s$ via proper randomness~$\vec y$. Most prominent Fiat-Shamir examples are DSA signatures and the new post-quantum standard Dilithium. In practice, DSA signatures have experienced fatal attacks via leakage of a few bits of the randomness~$\vec y$ per signature. Similar attacks now emerge for lattice-based signatures, such as Dilithium. We build on, improve and generalize the pioneering leakage attack on Dilithium by Liu, Zhou, Sun, Wang, Zhang, and Ming. {In theory}, their original attack can recover a 256-dimensional subkey of Dilithium-II (aka ML-DSA-44) from leakage in a single bit of $\vec{y}$ per signature, in any bit position $j \geq 6$. However, the memory requirement of their attack grows exponentially in the bit position $j$ of the leak. As a consequence, if the bit leak is in a high-order position, then their attack is infeasible. In our improved attack, we introduce a novel transformation, that allows us to get rid of the exponential memory requirement. Thereby, we make the attack feasible for \emph{all} bit positions $j \geq 6$. Furthermore, our novel transformation significantly reduces the number of required signatures in the attack. The attack applies more generally to all Fiat-Shamir-type lattice-based signatures. For a signature scheme based on module LWE over an $\ell$-dimensional module, the attack uses a 1-bit leak per signature to efficiently recover a $\frac{1}{\ell}$-fraction of the secret key. In the ring LWE setting, which can be seen as module LWE with $\ell = 1$, the attack thus recovers the whole key. For Dilithium-II, which uses $\ell = 4$, knowledge of a $\frac{1}{4}$-fraction of the 1024-dimensional secret key lets its security estimate drop significantly from $128$ to $84$ bits.

Lattice-based signatures like Dilithium (ML-DSA) prove knowl- edge of a secret key $s \in \mathbb{Z}_n$ by using Integer LWE (ILWE) samples $z = \langle \vec c, \vec s \rangle +y $, for some known hash value $c \in \mathbb{Z}_n$ of the message and unknown error $y$. Rejection sampling guarantees zero-knowledge, which makes the ILWE problem, that asks to recover s from many z’s, unsolvable. Side-channel attacks partially recover y, thereby obtaining more informative samples resulting in a—potentially tractable—ILWE problem. The standard method to solve the resulting problem is Ordinary Least Squares (OLS), which requires independence of $y$ from $\langle c, s \rangle$ —an assumption that is violated by zero-knowledge samples. We present efficient algorithms for a variant of the ILWE problem that was not addressed in prior work, which we coin Concealed ILWE (CILWE). In this variant, only a fraction of the ILWE samples is zero-knowledge. We call this fraction the concealment rate. This ILWE variant naturally occurs in side-channel attacks on lattice-based signatures. A case in point are profiling side-channel attacks on Dilithium implementations that classify whether $y = 0$. This gives rise to either zero-error ILWE samples $z = \langle c, s \rangle$ with $y = 0$ (in case of correct classification), or ordinary zero-knowledge ILWE samples (in case of misclassification). As we show, OLS is not practical for CILWE instances, as it requires a prohibitively large amount of samples for even small (under 10\%) concealment rates. A known integer linear programming-based approach can solve some CILWE instances, but suffers from two short-comings. First, it lacks provable efficiency guarantees, as ILP is NP-hard in the worst case. Second, it does not utilize small, independent error y samples, that could occur in addition to zero-knowledge samples. We introduce two statistical regression methods to cryptanalysis, Huber and Cauchy regression. They are both efficient and can handle instances with all three types of samples. At the same time, they are capable of handling high concealment rates, up to 90\% in practical experiments. While Huber regression comes with theoretically appealing correctness guarantees, Cauchy regression performs best in practice. We use this efficacy to execute a novel profiling attack against a masked Dilithium implementation. The resulting ILWE instances suffer from both concealment and small, independent errors. As such, neither OLS nor ILP can recover the secret key. Cauchy regression, however, allows us to recover the secret key in under two minutes for all NIST security levels.