International Association
for Cryptologic Research

Billions of dollars and countless GPU hours are currently spent on training Deep Neural Networks (DNNs) for a variety of tasks. Thus, it is essential to determine the difficulty of extracting all the parameters of such neural networks when given access to their black-box implementations. Many versions of this problem have been studied over the last 30 years, and the best current attack on ReLU-based deep neural networks was presented at Crypto'20 by Carlini, Jagielski, and Mironov. It resembles a differential chosen plaintext attack on a cryptosystem, which has a secret key embedded in its black-box implementation and requires a polynomial number of queries but an exponential amount of time (as a function of the number of neurons). In this paper, we improve this attack by developing several new techniques that enable us to extract with arbitrarily high precision all the real-valued parameters of a ReLU-based DNN using a polynomial number of queries \emph{and} a polynomial amount of time. We demonstrate its practical efficiency by applying it to a full-sized neural network for classifying the CIFAR10 dataset, which has 3072 inputs, 8 hidden layers with 256 neurons each, and about $1.2$ million neuronal parameters. An attack following the approach by Carlini et al.\ requires an exhaustive search over $2^{256}$ possibilities. Our attack replaces this with our new techniques, which require only 30 minutes on a 256-core computer.

Neural cryptanalysis is the study of cryptographic primitives through machine learning techniques. Following Gohr’s seminal paper at CRYPTO 2019, a focus has been placed on improving the accuracy of such distinguishers against specific primitives, using dedicated training schemes, in order to obtain better key recovery attacks based on machine learning. These distinguishers are highly specialized and not trivially applicable to other primitives. In this paper, we focus on the opposite problem: building a generic pipeline for neural cryptanalysis. Our tool is composed of two parts. The first part is an evolutionary algorithm for the search of good input differences for neural distinguishers. The second part is DBitNet, a neural distinguisher architecture agnostic to the structure of the cipher. We show that this fully automated pipeline is competitive with a highly specialized approach, in particular for SPECK32, and SIMON32. We provide new neural distinguishers for several primitives (XTEA, LEA, HIGHT, SIMON128, SPECK128) and improve over the state-of-the-art for PRESENT, KATAN, TEA and GIMLI.