Exploring Feature Selection Scenarios for Deep Learning-based Side-channel Analysis

One of the main promoted advantages deep learning in profiling sidechannel analysis is possibility skipping feature engineering process. Despite that, most recent publications consider selection as attacked interval from side-channel measurements pre-selected. This similar to worst-case security assumptions evaluations when random secret shares (e.g., mask shares) are known during phase: an evaluator can identify points ofinterest locations and efficiently trim trace interval. To broadly understand how impacts performance learning-based attacks, this paper investigates three different scenarios that could be realistically used practical evaluations. The range minimum possible number features (worst-case assumptions) whole available traces. Our results emphasize neural networks models show successful key recovery independently explored against first-order masked software implementations AES-128. First, we with optimal highly dependent on signal-to-noise ratio levels. Second, demonstrate attacking raw small also provides models, shortens gap between online (realistic) attacks. In all scenarios, hyperparameter search always indicates a model up eight hidden layers for MLPs CNNs, suggesting complex not required considered datasets. less than ten attack traces datasets at least one scenarios. Additionally, several cases, recover target single trace.