Enhancing Performance of Deep Learning Based Non-Profiled Side-Channel Attack Using Multi-Output and Transfer Learning

Multi-output deep learning based side-channel analysis (MO-DLSCA) has been introduced to improve the performance of non-profiled side-channel attacks (SCA) using deep learning. MO-DLSCA requires only one training process to reveal a subbyte of the secret key. In non-profiled SCA scenarios, attack techniques like Differential Power Analysis (DPA), Correlation Power Analysis (CPA), or MO-DLSCA must undergo the same process to attack different target devices, even if they share the same properties. This paper introduces a new SCA technique that leverages a multi-output deep learning neural network and transfer learning to enhance the performance of non-profiled SCA on targets with similar properties to the original target. Specifically, we demonstrate the advantage of our method in nonprofiled attacks over power traces collected from AVR XMEGA microcontrollers using ChipWhisperer. The experimental results show that our technique requires only $10 \%$ of the power traces compared to the first target to reveal the subbyte key in the second target. Additionally, the attack time for the second target is reduced significantly, approximately by a factor of 5.7, compared to the conventional MO-DLSCA technique.