A backpropagation extreme learning machine approach to fast training neural network-based side-channel attack

Abstract

In recent years, many side-channel attack (SCA) based on deep learning have emerged, making it possible to break protected encryption algorithms. How- ever, since the training of deep learning is based on back-propagation, a long training time is required. In deep learning SCA, because of the encryption algorithm, it is common to train multiple models based on the number of subkeys in one time attack, so the training time is multiplied as a drawback of DL- SCA. This work presented new Deep learning Side-channel Attack (DL-SCA) models that are based on Extreme Learning Machine (ELM). Unlike the conventional iterative backpropagation method, ELM is a fast learning algorithm that computes the trainable weights within a single iteration. Two models (Ensemble bpELM and CAE-ebpELM) are designed to perform SCA on AES with Boolean masking and desynchronization/jittering. The best models for both at- tack tasks can be trained 27× faster than MLP and 5× faster than CNN respectively. Verified and validated using ASCAD dataset, our models successfully recover all 16 subkeys using approximately 3K traces in the worst case scenario.