Adversarial Profiled Side-Channel Attack with Unsupervised Domain Adaptation

Deep learning-based methods have been proven to be very successful when applied to profiled side-channel attacks (SCA). In profiled SCA, the attacker gains knowledge about the target device by accessing a profiling device before the attack. However, most of the work does not consider the problem that it is difficult to obtain profiling traces that is fully consistent with the attack traces because the target device often employs more advanced masking countermeasures. Even if a small discrepancy between profiling traces and attack traces (regarded as domain discrepancy) may cause the attack failure. To address the problem of domain discrepancy, we propose a side-channel attack method based on the domain adversarial self-attention mechanism (DASA-SCA). Specifically, the method feeds the feature extractor through the bounded loss of the domain classifier when building the model in order to learn a hidden representation that is not only discriminative but also domain invariant. To enhance this domain-invariant representation, we introduce self-attention mechanism in the feature extractor to give higher weights to the significant parts of the leakage in order to improve the feature extraction. We implement three different deep learning models including CNN, ResNet, and LSTM in the DASA-SCA approach for both electromagnetic(EM) side-channel and power side-channel to demonstrate the scalability and generality of the framework. Finally, we validate the effectiveness of the DASA-SCA approach using the ASCAD dataset and DPA dataset. The experimental results demonstrate that our method can significantly reduce the impact of domain discrepancy caused by different countermeasures and enhance the side-channel attack capability.