Acnn: arbitrary trace attacks based on leakage area detection

Abstract

Deep Learning-based Side-Channel Analysis (DL-SCA) has emerged as a powerful method in the field of side-channel analysis. Current works on DL-SCA primarily rely on publicly available datasets, which typically consist of well-organized and well-aligned training and attack sets. However, this disregards the challenges faced in real-world attacks, where the attack traces are not well-aligned with the training traces as attackers have different levels of control over profiling and attack devices. A network that is capable of identifying areas of leakage and subsequently predicting the leaked values can bypass such difficulty. Therefore, we proposed Arbitrary Trace Attacks, which are placed under the flexible scenario that provides training traces and attack traces with arbitrary sizes. To implement such attacks, we present the Arbitrary Convolutional Neural Network (ACNN), which scans the input trace of arbitrary sizes for leakage area identification and leakage value prediction using a sliding window. Experimental evaluation is conducted on two datasets DPAv4.2 and ASCAD to verify the effectiveness of our approach on unprotected and masked implementation respectively. As a result, the target leakage areas are detected with a significant frequency and the key recovery performance is on par with state-of-the-art. Moreover, the trained model shows the potential for detecting leakage in a general context, that is, detecting leakage of key bytes other than the target one.