CL-SCCA: multi-byte contrastive learning for side-channel collision attacks on cryptosystems

Side-channel collision attacks (SCCA) are a classical method for recovering key differences by detecting collisions between intermediate values. Traditional SCCA methods usually rely on manually designed statistical distinguishers to measure the similarity between traces of different bytes, making it difficult to fully exploit the hidden shared leakage features among different byte traces under noise and nonlinear leakage. Deep learning-based SCCA improves the ability to extract leakage features, but existing methods are usually trained around a single byte and therefore cannot fully utilize the richer leakage distribution provided by multi-byte traces. To address this issue, this paper proposed a Contrastive Learning-based SCCA method, called CL-SCCA. CL-SCCA uses traces corresponding to multiple byte operations for pretraining, enabling the encoder to capture similar leakage features from the multi-byte leakage distribution, thereby improving key-difference recovery. Furthermore, this paper proposed a byte-weighted contrastive loss function to reduce the influence of byte-source differences on model training during pretraining. Across datasets covering different cryptographic primitives and hardware platforms, the proposed method consistently achieved better attack performance, with relative improvements in recovery success rate of approximately 14.88%, 77.73%, and 65.20% on ASCAD_F, ASCAD_R, and DPA Contest v4, respectively.