Non-profiling Time-Frequency Analysis based Correlation Optimization with Deep Learning

Correlation Optimization with Deep Learning Analysis (CO-DLA) is a state-of-the-art non-profiling side-channel analysis based on Deep Learning. In this paper, we proposes a novel Wavelet Scattering Transform-based CO-DLA (WST-CO-DLA). WST-CO-DLA utilizes the non-training-required Wavelet Scattering Transform (WST) to extract time-frequency features for CO-DLA, thus offering three main contributions. Firstly, WST eliminates the need for training-required convolution blocks in CNN-based CO-DLA or reduces the number of linear layers in MLP-based CO-DLA, resulting in faster execution time as compared to existing CO-DLA methods. WST-CO-DLA performs 28% and 60% faster than the reported MLP-based CO-DLA and CNN-based CO-DLA, respectively, on the ASCAD datasets. Secondly, WST transforms the 1-D time domain Physical Leakage Information (PLI) measurements to a 2-D time-frequency domain, providing rich time-frequency features for WST-CO-DLA to learn from. Experiments show that WST-CO-DLA requires 50% and 75% fewer PLI measurements to reveal the secret key from the aligned and misaligned ASCAD datasets, respectively, compared to reported CO-DLA methods. Thirdly, we implement a Short Time Fourier Transform-based CO-DLA (STFT-CO-DLA) and benchmark WST-CO-DLA against STFT-CO-DLA. Experimental results show that WST-CO-DLA requires up to 95.83% fewer measurements than STFT-CO-DLA to extract the secret key from the ASCAD datasets.