Exploring Multi-task Learning in the Context of Masked AES Implementations

Deep learning is very efficient at breaking masked implementations even when the attacker does not assume knowledge of the masks. However, recent works pointed out a significant challenge: overcoming the initial learning plateau. This paper discusses the advantages of multi-task learning to break through the initial plateau consistently. We investigate different ways of applying multi-task learning against masked AES implementations (via the ASCAD-r, ASCAD-v2, and CHESCTF-2023 datasets) under the assumption that the attacker cannot access masks during training. We offer evidence that multi-task learning significantly increases the consistency of convergence and performance of deep neural networks. Our work provides a wide range of experiments to understand the benefits of multi-task strategies over the current single-task state-of-the-art. Furthermore, such strategies achieve novel milestones against protected implementations as we propose models that defeat all masks of the affine masking on ASCAD-v2 for the first time.