Convolutional Neural Network Based Side-Channel Attacks with Customized Filters

Deep learning is progressively gaining attention as a powerful tool for conducting profiling side-channel attacks. In particular, convolutional neural network (CNN) is one of the mostly employed learning techniques in the context of side-channel analysis. The first layer of a standard CNN always performs a set of convolutions between the input and some finite impulse response filters. In this work, we substitute the standard filter by a customized filter borrowed from the domain of speaker recognition due to the resemblance between the power traces and speech signals. In contrast to standard filters, the new filter only depends on parameters with a clear physical meaning, where only low and high cutoff frequencies are learned from the training data. Experimental results obtained from public datasets show that the side-channel attacks based on CNNs equipped with this new filter are more effective and robust than attacks based on standard CNNs. The results of this work open new perspective and encourage further research on the effect of the filters of the CNN-based side-channel attacks.