Efficient Adversarial Learning-Based Profiled Side-Channel Analysis via Source Encoder Freezing and Randomized Multilinear Mapping

Deep learning-based side-channel analysis has been proven to be a powerful method for attacking cryptographic devices. However, it faces a portability problem due to differences between the profiling device and the target device. At DAC 2022, Cao et al. proposed an adversarial learning-based profiling side-channel analysis (AL-PA) approach to mitigate this issue. Nevertheless, this method has not been systematically analyzed under challenging scenarios such as low signal-to-noise ratios and protection countermeasures. To address this gap, this paper investigates the performance limitations of AL-PA under challenging scenarios. We identify that fully sharing encoder parameters across domains restricts the model’s adaptability when the source and target distributions diverge significantly. To overcome this limitation, we propose a training strategy that freezes the source encoder parameters, thereby improving the stability and adaptability of adversarial training. Furthermore, a randomized multilinear mapping is introduced to replace the original conditional outer product, which reduces the number of discriminator parameters while preserving classification performance. We validate the proposed method on the SAKURA-AES and XMEGA-SM4 datasets. Experimental results show that our approach reduces discriminator parameters by at least 96.87%, and lowers the number of required attack traces by 37.81% compared to AL-PA.