Metaheuristic-driven neural architecture search for deep learning-based side-channel analysis

Deep Learning (DL) has proven highly effective in Side-Channel Analysis (SCA), especially against secure devices like smart cards. However, designing efficient DL models remains time-consuming and often unsuitable for new or unknown systems, as performance heavily depends on hyperparameters and architecture choices. To address this, we investigate the integration of metaheuristic algorithms into Neural Architecture Search (NAS) for SCA. We study two metaheuristic classes-population-based metaheuristics (Genetic Algorithm (GA) and Particle Swarm Optimization (PSO)) and single-solution-based metaheuristics (Simulated Annealing (SA) and Tabu Search (TS))-to automatically design and optimize DL models. Two search spaces are modeled: a smaller one for Multilayer Perceptrons (MLPs) and a larger, more complex one for Convolutional Neural Networks (CNNs), enabling evaluation under varying search complexities. Experiments on the ASCAD (Fixed Key) and CHES CTF datasets show that metaheuristic-NAS consistently outperforms traditional optimization techniques such as Random Search (RS) and Bayesian Optimization (BO). In smaller spaces, SA offers competitive results with lower execution time, while Tabu Search finds good architectures but is slower. Using the Guessing Entropy (GE) metric, in the MLP search space the SA approach achieves GE = 1 in  ≈  60 traces. In the CNN space, top configurations reach GE = 1 in  ≈  200 traces for ID leakage and  ≈  420 for the more challenging Hamming Weight leakage. Based on our experimental evaluation, metaheuristic-NAS hybrids are confirmed to be efficient and practical tools for automating DL model design in SCAs.