Domain Adaptation for Side-Channel Analysis Using Deep Reconstruction-Classification Networks

Abstract

Side-channel analysis (SCA) exploits physical leakages to evaluate cryptographic system security, but traditional methods struggle with cross-device attacks due to hardware variations. This paper introduces a domain adaptation approach using Deep Reconstruction-Classification Networks (DRCN) and Maximum Mean Discrepancy (MMD) loss, enabling knowledge transfer without labeled data. By aligning feature distributions between source and target devices, the DRCN method improves cross-device attack performance. Experiments on datasets, including TinyAES-128 electromagnetic traces and XMEGA data, show significant enhancements. The approach reduces the number of power traces required for key recovery and lowers Guessing Entropy (GE) in cross-device scenarios, increasing attack success rates. In summary, the DRCN-based domain adaptation method offers an effective solution for improving cross-device SCA, demonstrating practical value and efficiency.