Toward practical autoencoder-based side-channel analysis evaluations

Abstract

In the field of side-channel analysis, profiled attacks are one of the most powerful types of attacks. Nevertheless, major issues with profiled attacks are in sensitivity to noise and the high-dimensional nature of the signals used for training, generating a less efficient classifier to conduct the attack phase. Consequently, evaluating the security of cryptographic implementation in hardware devices like IoT becomes more complex as side-channel analysis evaluation easily falls into false-positive results. This paper assesses the efficacy of applying a feature reduction process to deal with high-dimensional signals. We propose a practical procedure to conduct feature reduction using autoencoders for profiled side-channel leakage evaluations. Two autoencoder architectures are compared while performing feature reduction showing that our proposed architecture keeps most of the relevant information. Our proposal is tested on the ASCAD random key database with a high desynchronization value and produced results that outperform other state-of-the-art techniques. The guessing entropy value converges to 1 after around 500 leakage traces.