Interpreting Information of Deep Neural Networks for Profiled Side Channel Analysis

Abstract

Security has become ever more important in today's quickly growing digital world as the number of digital assets has quickly grown. Our thesis focuses on devices that compute a secure cryptographic operation such that information can be communicated or authenticated. The attack vector utilized is known as Profiled Side-Channel Analysis (SCA) which aims at extracting a cryptographic key from a device through unintended behavior expressed through power monitoring or electromagnetic radiation. Profiled SCA attacks assume the most powerful adversary and therefore allows us to make a sound security assessment of a device in this setting. Our utilized profiling technique includes deep neural networks such as the multi-layer perceptron and the convolutional neural network. As this adds a layer of complexity to our assessment, we must understand how the properties of the network consolidate our security assessment. Previous research has shown that classical neural network metrics such as accuracy does not correlate to how successful or efficient a side-channel analysis is, therefore, we have proposed a mutual information metric. We measure mutual information across each layer in the neural network such that the behavior of each layer in interpreting how each layer is benefiting our classification. We investigate if the mutual information measure can be used to make a beneficial architectural distinction of the neural network for our side-channel analysis problem. Finally, we show there is a relationship between the mutual information and the guessing entropy for our side-channel attack and that it can be used to confirm that the chosen model is fully optimized for the side-channel problem.