International Association
for Cryptologic Research

The machine learning problem of extracting neural network parameters has been proposed for nearly three decades. Functionally equivalent extraction is a crucial goal for research on this problem. When the adversary has access to the raw output of neural networks, various attacks, including those presented at CRYPTO 2020 and EUROCRYPT 2024, have successfully achieved this goal. However, this goal is not achieved when neural networks operate under a hard-label setting where the raw output is inaccessible. In this paper, we propose the first attack that theoretically achieves functionally equivalent extraction under the hard-label setting, which applies to ReLU neural networks. The effectiveness of our attack is validated through practical experiments on a wide range of ReLU neural networks, including neural networks trained on two real benchmarking datasets (MNIST, CIFAR10) widely used in computer vision. For a neural network consisting of $10^5$ parameters, our attack only requires several hours on a single core.

The differential-linear attack is one of the most effective attacks against ARX ciphers. However, two technical problems are preventing it from being more effective and having more applications: (1) there is no efficient method to search for good differential-linear approximations. Existing methods either have many constraints or are currently inefficient. (2) partitioning technique has great potential to reduce the time complexity of the key-recovery attack, but there is no general tool to construct partitions for ARX ciphers. In this work, we step forward in solving the two problems. First, we propose a novel idea for generating new good differential-linear approximations from known ones, based on which new searching algorithms are designed. Second, we propose a general tool named partition tree, for constructing partitions for ARX ciphers. Based on these new techniques, we present better attacks for two ISO/IEC standards, i.e., LEA and Speck. For LEA, we present the first 17-round distinguisher which is 1 round longer than the previous best distinguisher. Furthermore, we present the first key recovery attacks on 17-round LEA-128, 18-round LEA-192, and 18-round LEA-256, which attack 3, 4, and 3 rounds more than the previous best attacks. For Speck, we find better differential-linear distinguishers for Speck48 and Speck64. The first differential-linear distinguishers for Speck96 and Speck128 are also presented.