CNN-Based Fault Classification in Side-Channel Power Traces with Robust Preprocessing and Regularization

Abstract

Side-channel attacks are carried out by observing physical leakages such as power, timing, or electromagnetic emissions, and these leakages often reveal information about the internal operations of a cryptographic device. In this work, synthetic power-trace data with controlled fault patterns was generated so that the behavior of different machine learning models could be studied under a variety of fault conditions. Several types of artificial disturbances were injected into the traces using a structured simulation process, allowing both normal and faulty examples to be created at scale. The generated dataset was then used to evaluate multiple models, beginning with an autoencoder for feature compression followed by a Random Forest classifier on the encoded outputs. A deeper CNN model was later applied to learn features directly from raw traces, and this approach produced the best results. The CNN was also extended to a multi-class setting, and various optimizers were tested to check their influence on training stability. Overall, the models based on convolutional networks were found to perform the most reliably, indicating that deep feature extraction works better for high-dimensional side-channel data.