Paper 2020/1258

TranSCA: Cross-Family Profiled Side-Channel Attacks using Transfer Learning on Deep Neural Networks

Dhruv Thapar, Manaar Alam, and Debdeep Mukhopadhyay

Abstract

Side-channel analysis (SCA) utilizing the power consumption of a device has proved to be an efficient technique for recovering secret keys exploiting the implementation vulnerability of mathematically secure cryptographic algorithms. Recently, Deep Learning-based profiled SCA (DL-SCA) has gained popularity, where an adversary trains a deep learning model using profiled traces obtained from a dummy device (a device that is similar to the target device) and uses the trained model to retrieve the secret key from the target device. \emph{However, for efficient key recovery from the target device, training of such a model requires a large number of profiled traces from the dummy device and extensive training time}. In this paper, we propose \emph{TranSCA}, a new DL-SCA strategy that tries to address the issue. \emph{TranSCA} works in three steps -- an adversary (1) performs a one-time training of a \emph{base model} using profiled traces from \emph{any} device, (2) fine-tunes the parameters of the \emph{base model} using significantly less profiled traces from a dummy device with the aid of \emph{transfer learning} strategy in lesser time than training from scratch, and (3) uses the fine-tuned model to attack the target device. We validate \emph{TranSCA} on simulated power traces created to represent different FPGA families. Experimental results show that the transfer learning strategy makes it possible to attack a new device from the knowledge of another device even if the new device belongs to a different family. Also, \emph{TranSCA} requires very few power traces from the dummy device compared to when applying DL-SCA without any previous knowledge.