Advanced Side-Channel Profiling Attacks with Deep Neural Networks: A Hill Climbing Approach

Abstract

Deep learning methods have significantly advanced profiling side-channel attacks. Finding the optimal set of hyperparameters for these models remains challenging. Effective hyperparameter optimization is crucial for training accurate neural networks. In this work, we introduce a novel hill climbing optimization algorithm that is specifically designed for deep learning in profiled side-channel analysis. This algorithm iteratively explores hyperparameter space using gradient-based techniques to make precise, localized adjustments. By incorporating performance feedback at each iteration, our approach efficiently converges on optimal hyperparameters, surpassing traditional Random Search methods. Extensive experiments—covering protected implementations, leakage models, and various neural network architectures—demonstrate that our hill climbing method consistently achieves superior performance in over 80% of test cases, predicting the secret key with fewer attack traces and outperforming both Random Search and state-of-the-art techniques.