Convolutional Neural Network Based Side-Channel Attacks in Time-Frequency Representations

Profiled attacks play a fundamental role in the evaluation of cryptographic implementation worst-case security. For the past sixteen years, great efforts have been paid to develop profiled attacks from Template Attacks to deep learning based attacks. However, most attacks are performed in time domain – may lose frequency domain information. In this paper, to utilize leakage information more effectively, we propose a novel deep learning based side-channel attack in time-frequency representations. By exploiting time-frequency patterns and extracting high level key-related features in spectrograms simultaneously, we aim to maximize the potential of convolutional neural networks in profiled attacks. Firstly, an effective network architecture is deployed to perform successful attacks. Secondly, some critical parameters in spectrogram are studied for better training the network. Moreover, we compare Template Attacks and CNN-based attacks in both time and time-frequency domain with public datasets. The heuristic results in these experiments provide a new perspective that CNN-based attacks in spectrograms give a very feasible option to the state-of-the-art profiled attacks.