Year: 2023
Type: conference-paper
Abstract: This paper provides a novel perspective on improving the efficiency of side-channel analysis by applying two deep feature loss functions: Soft Nearest Neighbor (SoftNN) and Center loss. By leveraging these loss functions during the deep neural networks (DNNs) training phase, our study illuminates how profiling attacks can be more powerful. Deep feature loss functions incorporate the outputs from the DNN's intermediate layers into their computations, which reduces the distance between similar data points. As such, these techniques enhance the DNN's ability to generate more precise and meaningful representations, thereby improving its discriminative power. This paper presents empirical evidence illustrating the effectiveness of SoftNN and Center loss in strengthening DNN-based side-channel attacks. For instance, when using Center loss together with the focal loss ratio (FLR), it requires the least number of traces to break the ASCADf dataset. On the other hand, applying SoftNN with FLR successfully recovers the key for the ASCADr dataset with the least traces. The insights presented in this study can act as a baseline for more advanced investigations into the utility of such loss functions in deep learning-based side-channel analysis.
Authors: Trevor Yap, Stjepan Picek, Shivam Bhasin
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4388958832