부산대학교 도서관

Deep learning-driven side-channel analysis (SCA) is a promising approach to side-channel analytic profiling. Recent studies have shown that neural networks can successfully attack defended targets, even with a small number of attack traces. However, developing neural networks requires fine-tuning hyperparameters, which is challenging and time-consuming, especially for complex neural networks. This study proposes an AutoSCA framework that uses Bayesian optimization to automate deep learning hyperparameter tuning for SCA. The framework is implemented using two popular neural network architectures: the multi-layer perceptron (MLP) and convolutional neural network (CNN). The AutoSCA framework improves deep learning performance and side-channel measurements, which has potential applications in 6G communication-based mobile devices. The framework was trained and evaluated using the ASCAD and CHES CTF datasets. The experimental results showed that the CNN-based AutoSCA outperformed the MLP-based AutoSCA and other state-of-the-art models, in terms of low time complexity and higher accuracy. Results suggest that Bayesian optimization is effective regardless of the dataset, neural network architecture, or type of leaky prototype in defeating contemporary attacks. Applying deep learning optimization against side-channel attacks in consumer electronics can significantly enhance the security of user data and privacy in an increasingly connected.