Wavelet Coefficients Based Generative Adversarial Networks for Side-Channel Attack Preprocessing

Wavelet Coefficients Based Generative Adversarial Networks for Side-Channel Attack Preprocessing

연구 분야: Cryptography

학회: Signal, Image and Video Processing

As cryptographic systems evolve, increasingly sophisticated countermeasures have been developed that significantly challenge the efficacy of side-channel attacks by introducing distortions and noise into leakage traces. Current denoising methods often fail to effectively mitigate environmental noise and interference induced by these countermeasures while preserving essential leakage patterns. This study introduces a novel framework that integrates wavelet coefficients into a generative adversarial network (GAN) for the advanced preprocessing of side-channel traces. Utilizing the discrete wavelet transform, signals are decomposed into high-frequency and low-frequency subbands, which facilitates frequency-aware noise suppression by employing a conditional GAN architecture. This technique ensures the preservation and recovery of discriminative leakage features across various spectral domains, while maintaining the fidelity of the traces. When evaluated using the ASCAD dataset, our framework significantly outperforms both unprocessed and mean-filtered traces, achieving a signal-to-noise ratio (SNR) of 7.87 and a Pearson correlation coefficient (PCC) of 0.198. Additionally, the preprocessing reduces the number of traces required for successful key recovery by 43% (from 191 to 108 traces), thereby enhancing the efficiency of the attack. Experimental results confirm the framework’s capability to counteract masking countermeasures and ambient noise without reducing the number of traces, addressing a critical limitation of conventional methods. This work contributes novel preprocessing methodologies for side-channel attacks and underscores the potential of frequency-domain deep learning techniques for robust leakage extraction in practical cryptographic systems.