Efficient Side-Channel Attack through Balanced Labels Compression and Variational Autoencoder

Recently, side-channel attacks based on deep learning (DLSCAs) have attracted much attention. Many works have improved the performance of DLSCAs by designing advanced neural network architectures and training strategies. There are few studies on leakage models for DLSCAs. Existing researches usually utilize the intermediate value Hamming weight (HW) and the intermediate value itself (ID) as leakage models. Training a classifier with good performance is challenging due to the many label classes in the ID leakage model. The HW leakage model can significantly reduce the number of labels, but it will cause samples imbalance. In this paper, we propose a new DLSCA leakage model, named Balanced Labels Compression (BLC). We consider dividing sensitive intermediate values with same lowest bits into same class to obtain balanced labels. Then, we train a classifier using the compressed BLC labels and profiling energy traces. At the attack phase, the probability distribution of BLC labels is extended to the probability distribution of sensitive intermediate values. We conduct extensive comparison experiments with HW, ID, and BLC leakage models under the two scenarios of sufficient and insufficient profiling energy traces. Further, we exploit VAE to improve attack performance when energy traces are insufficient. Experimental results show that VAE-based data augmentation can significantly reduce required energy traces.