Cover to Uncover: Comprehensive Study of Occlusion in DL-based SCA

Deep Learning-based side-channel analysis (SCA) has gained a lot of popularity and represents a powerful addition to classical methods. A significant portion of this field's research focuses on new architectures or hyper-parameters to improve performance. The found models are typically assessed by their prediction capabilities, but which sample points of each dataset in combination with the found architecture were decisive is rarely discussed. We extend the work of Schamberger et al. by focusing on occlusion as a lightweight explainability method. We present for the first time an in-depth investigation of the 2018 CHES CTF AES dataset based on the Deep Neural Network (DNN) results of Gohr et al. We propose a new variant, uniform random occlusion, which excels in combination with small occlusion windows. We use random 1-occlusion to identify points of interests (PoIs) in the dataset and compare those results with PoIs generated with the saliency maps and with a classical correlation-based PoI search method. For the first time in SCA, we apply occlusion during the training phase of a DNN. We show that in this case, the model partially extracts information from previously unimportant PoIs. Our findings suggest that a combination of DNNs trained on occluded data may enhance prediction accuracy.