Attention - Based Non-Profiled Side-Channel Attack

Profiled attacks play a fundamental role in evaluating the security of cryptographic algorithm implementations in the worst case. In recent years, deep learning has been introduced as a new powerful alternative to achieve these attacks. Many works have demonstrated that deep learning could solve the desynchronization issue and break the masking countermeasures. From the early years when researchers explore the machine learning methods, there is a road map that the profiled attacks could be converted to non-profiled ones by on-the-fly strategy. In TCHES 2019, the Differential Deep Learning Analysis (DDLA) combining the deep learning and non-profiled attacks by on-the-fly strategy was proposed. The DDLA mainly uses the sensitivity analysis as the attack metric. In this paper, we propose a new metric, attention mechanism, to conduct deep learning-based non-profiled attacks. As a part of the feed-forward structure, when attention is used as the metric, it does not need additional gradient computations like the DDLA. Meanwhile, the attention probabilities also indicate where the intermediate leaks, not on the input traces but on the abstract feature vectors, which gives the attacker more insights into how the network makes the decision. Finally, we test our attention-based non-profiled attack on practical traces. The results show that it works well on desynchronized traces and masked implementations.