Impact of Data Pre-Processing Techniques on Deep Learning Based Power Attacks

Abstract

Power-based side channel attacks (SCAs) are recognized as a powerful type of hardware attacks. Recently, attacks based on deep learning (DL) neural networks have become popular due to their high efficiency. However, even these attacks face problems when sophisticated countermeasures exist. Pre-processing the input data is an effective way to improve the performance of such neural networks. Currently, only limited research has focused on exploring pre-processing techniques for DL-based attacks. In this paper, we propose to the best of our knowledge for the first time the usage of data transformation, data concatenation and stacked auto-encoder (encoder only) as pre-processing methods. Thereafter, we compare them with the existing techniques, namely data augmentation and stacked auto-encoder techniques. Our results show that the data transformation technique achieves the best results from the evaluated methods; it improves the validation accuracy from 75% to 95% and 23% to 26% for the RSA and AES implementations, respectively.