Switch-T: A novel multi-task deep-learning network for cross-device side-channel attack

Abstract

Side-Channel Analysis has become a realistic threat to cryptographic implementations, particularly with advances in deep-learning techniques. A well-trained neural network can typically make the attack several orders of magnitude more efficient than conventional signal processing approaches. However, like all profiled methods, most existing deep-learning SCAs frameworks require adversaries to develop dedicated models for the specific target device, which complicates the execution of these attacks. In this paper, we propose a Transformer-based neural network, called Switch-T, for multi-task attacks. By collaboratively employing the Elastic Weight Consolidation (EWC) mechanism with a multi-task structure, the model is feasible to learn sensitive data-dependent features of power and EM traces from devices with different core architectures and PCB layout. We experimentally show that the Switch-T model can effectively compromise different implementations of AES. Furthermore, we investigate to which extent the training order of profiling devices can affect the attack efficiency of the model and discuss the impact of hyper-parameter settings in the EWC mechanism.