Hide and Seek: Using Occlusion Techniques for Side-Channel Leakage Attribution in CNNs

Deep learning-based side-channel analysis has gained popularity due to its relaxed feature engineering effort in contrast to classical profiled side-channel analysis approaches. This however comes at the cost of a reduced explainability of attack results. In this work we propose occlusion techniques for neural network attribution that allow the identification of points-of-interest related to side-channel leakage used by the networks to defeat masking countermeasures. We evaluate results for both ASCAD databases and are able to identify occlusion parameters that are suitable in the side-channel context. We reason that due to side-channel measurement characteristics multiple adjacent samples have to be occluded at once, which has not been considered in related work. In addition, with our higher-order occlusion we are able to identify leakage combinations that are exploited by a network in order to mount a higher-order attack. Using our methods we are able to show that networks actually utilize varying leakage characteristics observable for different key bytes of the ASCAD databases. This work shows that occlusion is a viable addition to established gradient-based attribution methods.     «

Deep learning-based side-channel analysis has gained popularity due to its relaxed feature engineering effort in contrast to classical profiled side-channel analysis approaches. This however comes at the cost of a reduced explainability of attack results. In this work we propose occlusion techniques for neural network attribution that allow the identification of points-of-interest related to side-channel leakage used by the networks to defeat masking countermeasures. We evaluate results for both...     »