Hybrid genetic algorithm and deep learning techniques for advanced side-channel attacks.

In recent years, deep learning-based profiling methods have significantly advanced side-channel analysis, yielding promising results. A critical challenge in training effective neural network models lies in hyperparameter optimization. This research introduces a genetic algorithm (GA) framework that efficiently navigates complex hyperparameter search spaces, overcoming limitations of conventional methods: grid search's poor scalability and Bayesian optimization's challenges with high-dimensional spaces. The GA leverages evolutionary strategies to explore non-differentiable, multimodal optimization landscapes, systematically identifying configurations that maximize model performance. Our evaluation on protected AES implementations demonstrates the framework's effectiveness. Experimental results show the GA-based approach achieves 100% key recovery accuracy across test cases, significantly outperforming random search baselines (70% accuracy). In comprehensive comparisons against Bayesian optimization, reinforcement learning, and tree-structured Parzen estimators, the GA solution achieved top performance in 25% of test cases and ranked second overall. These findings validate genetic algorithms as a robust alternative for optimizing side-channel attack models, offering both scalability and consistent performance across diverse attack scenarios while advancing the state of cryptographic security assessment.