Enhancing deep learning-based side-channel analysis using feature engineering in a fully simulated IoT system

The increasing integration of cloud and embedded systems has made security more critical. Despite efforts to implement countermeasures against attacks, new threats have constantly emerged. Deep learning (DL) is most notable for side-channel disassembly attacks that expose cloud-to-things operations. This underscores the need to develop effective tools to test a system's robustness against such attacks. In this study, we developed a robust instruction-level side-channel disassembler for hiding countermeasures in a fully simulated embedded system. We investigated the effect of a moving-window-based feature engineering technique using statistical methods on the performance of side-channel disassembly attacks orchestrated via DL models. In addition, we propose a moving log-transformed temporal integration feature that enhances the performance of DL models for detecting and inferencing tasks. The created dataset was applied for two DL tasks: detecting hiding countermeasures and inferring assembly instructions. Using our feature engineering method, we found that the artificial neural network (ANN) showed an accuracy of 98.81% for hiding countermeasure detection, and the gated recurrent unit (GRU) model inferred the assembly sequence with 98.7% accuracy. These results highlight the need for advanced hardware- and software-level security measures to prevent side-channel attacks on embedded devices as potential vulnerabilities in the cloud infrastructure.