Multi-label Deep Learning based Side Channel Attack

In this paper, a novel side channel attack method with multi-label deep learning is proposed, and it surpasses the state-of-the-art result in ASCAD benchmark dataset. Our experimental results show that an ingenious modification of the output layer of a neural network can bring several times improvement of attack performance than the original model, even dozens of times when dealing with unaligned traces. In fact, the masked AES implementation of ASCAD only requires about 150, 250 and 350 traces to break respectively, when traces are manually desynchronized with our multi-label classification model. Our model can be considered as an ensemble learning method using the integration of monobit models based on the independence assumption of bits in the intermediate value.