Effective Deep Learning-based Side-Channel Analyses Against ASCAD

Side-channel analysis (SCA) based on deep learning (DL) techniques have the benefit that they can disclose the secret key of protected block ciphers without preprocessing. But the size of convolutional neural network (CNN) architecture is so large that the training process is too time-consuming and the required number of traces for recovering secret key is too much. In this paper, we apply heatmap and SNR to reduce the number of parameters of our CNN architecture to 5,269,568 (i.e., one to tenth to the CNNbest). We also combine our CNN architecture with the multi-label classification and the transfer learning techniques to reduce the number of required traces. Consequently, two new CNN architectures are presented and validated on the public ASCAD dataset. The execution time of the training process approximates 15 minutes on average. The first CNN architecture can recover a key byte with only 30 synchronized traces. Combined with the transfer learning technique, the second CNN architecture requires 141 and 171 traces respectively in two different desynchronization cases. To our knowledge, for both synchronization and desynchronization cases, our analysis methods need the smallest amount of traces to extract a key byte.