Side channels are electronic device interfaces that can unintentionally reveal information about the device’s state or processing. They have become a source of vulnerabilities in cryptographic algorithms, leading to side-channel attacks that can compromise sensitive information. Deep learning models have emerged as effective tools for side-channel evaluation due to their ability to interpret side-channel information. However, designing an efficient deep learning network for side-channel attack evaluation requires consideration of various factors, such as network architecture and hyperparameters selection. Furthermore, a single deep learning model cannot evaluate multiple cryptographic implementations, which limits its practicality.

To address these challenges, this thesis proposes new processes and methodologies to improve the training process of deep learning models and modify their architecture for side-channel analysis evaluation. The research focuses on integrating Six Sigma methodology, early stopping framework, feature reduction, and transfer learning approaches to enhance the training process. Additionally, new ways to design the architecture of the deep learning model are proposed. The aim is to improve the efficiency of side-channel evaluation by designing deep learning models that can evaluate multiple cryptographic implementations.