Insung Kim | Searcharxiv

arXiv2026

MIFA: An MILP-based Framework for Improving Differential Fault Attacks

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

cs.CR↗

arXiv2026

Provably Learning Multi-Head Attention with Queries

We study the problem of learning multi-head softmax attention from black-box input-output access. The learner may query arbitrary real-valued token sequences and observe only the scalar output at the final token. Recent work gives an algorithm using $O(d^2)$ value queries to recover the single-head parameters $(W,v)$. For multiple heads, the same work establishes identifiability under the assumption that the heads occupy pairwise orthogonal subspaces. Applying the single-head recovery algorithm separately to the heads additionally requires bases for these subspaces to be known. We recover a canonical representation by merging heads with the same $W_h$, summing their corresponding $v_h$, and discarding a merged head when this sum is zero, without these subspace assumptions. By varying the number of copies of a token, our algorithm obtains samples of a rational function whose interpolation separates the canonical heads. Additional queries formed by adding selected token vectors then match the same head across different queries. When the oracle outputs and all subsequent computations are exact, the learner chooses its query vectors at random and recovers the canonical pairs $\{(W_h,v_h):h\in[H]\}$ up to permutation with probability one. When $H$ is known, it uses exactly $4Hd^2-2H+1$ value queries of maximum length $2H+1$. If only a known upper bound $H_0$ is available, the algorithm uses $4H_0d^2-2H_0+1$ value queries of maximum length $2H_0+1$. For approximate oracle outputs, we give conditions under which the parameter error is at most a model- and query-dependent constant multiple of the output error. Finally, we extend our result to a one-layer Transformer with multi-head attention followed by a bias-free ReLU feed-forward network. Under additional conditions, we recover a functionally equivalent Transformer without relying on a separate algorithm for learning the feed-forward network.

cs.LG↗