Improving Location-Based Thermal Emission Side-Channel Analysis Using Iterative Transfer Learning

Kai Zhang, Tun-Chieh Lou, Chung-Che Wang, Jyh-Shing Roger Jang, Henian Li, Lang Lin, Norman Chang

Abstract: This study applies machine learning to side-channel attacks and proposes an iterative transfer learning method for deep learning models. This study leverages the similarity in training patterns across bytes by first training on a single byte and then using the resulting model as a pretrained foundation for the remaining bytes. This approach enables effective model training with smaller amounts of data while reducing the measurement-to-disclosure (MTD, i.e., the minimum number of traces needed for successful key recovery) in the attack phase. With sufficient data, iterative transfer learning reduces MTD from 55 to 54 using MLP and from 125 to 83 using CNN. Even under limited data conditions, it successfully breaks AES-128 while reducing training samples from 13,600 to 2,000, achieving an average MTD of 635, whereas traditional methods fail. Experimental results demonstrate that the iterative transfer learning approach addresses the persistent data scarcity challenge in deep learning, significantly expanding the applicability of deep learning methods in side-channel attack scenarios.