RFA-SCA: Robust Feature Alignment for Side-Channel Analysis via Multi-Order Moment Alignment

The effectiveness of profiling deep learning side-channel attacks relies on the assumption that training and attack data follow the same distribution. However, when the profiling device differs from the target device, process-voltage-temperature (PVT) variations and clock jitter countermeasures cause distribution shifts in power traces, rendering models trained on the source device ineffective on the target. Existing domain adaptation methods typically rely on a single distributional constraint without jointly constraining kernel mean embeddings and covariance structure, thus limiting their effectiveness against strong defenses such as clock jitter. We propose Robust Feature Alignment for Side-Channel Analysis (RFA-SCA), an unsupervised domain adaptation framework that aligns source and target feature distributions without requiring the target-device key during training. RFA-SCA combines three loss components. MMD loss reduces distributional discrepancy via kernel mean embeddings in a reproducing kernel Hilbert space (RKHS), CORAL loss regularizes covariance-level feature discrepancies, and conditional entropy loss optimizes decision boundaries in the target domain. In randomized attack trials using a finite number of target-domain traces on four benchmark datasets (ASCAD, CHES CTF 2018, SAKURA-G, XMEGA), RFA-SCA achieves 100% key recovery across all six cross-device and countermeasure scenarios. Furthermore, RFA-SCA reduces the number of traces required for successful attacks by 14%–33% compared to the best baseline in the ASCAD Desync, ASCAD Gaussian Noise, and SAKURA-G scenarios. A systematic ablation study over seven loss-component variants across all six scenarios further supports the role and complementarity of the three components.