The Forgotten Hyperparameter: Introducing Dilated Convolution for Boosting CNN-Based Side-Channel Attacks

摘要

In the evaluation of side-channel resilience, convolutional neural network-based techniques have been proved to be very effective, even in the presence of countermeasures. This work is introducing the use of dilated convolution in the context of profiling side-channel attacks. We show that the convolutional neural network that uses dilated convolution increases its performance by taking advantage of the leakage distributed through scattered points in leakage traces. We have validated the feasibility of the proposal by comparing it with the state-of-the-art approach. We have conducted experiments using ASCAD (with random key), and as a result the guessing entropy of the attack converges to zero for around 550 synchronized traces and for 3 000 desynchronised traces. In both groups of experiments, we have used the same architecture to train the model, changing just dilatation rate and kernel length, which indicates a reduction of the complexity in the deep learning model.