Strive to Fail: Deep Learning-based Side-channel Analysis for Evaluators

Digital devices are now deeply embedded in modern life. These devices process sensitive information, including personal data, financial records, and data related to critical infrastructure. Cryptography is therefore a fundamental component of digital security, providing confidentiality, integrity, authentication, key exchange, and digital signatures.

Although cryptographic algorithms are designed to be mathematically secure, their physical implementations can introduce vulnerabilities. When cryptographic algorithms run on hardware, devices unintentionally leak information through side channels such as power consumption, electromagnetic radiation, and timing behavior. These leakages can be exploited through side-channel analysis to recover secret information, including cryptographic keys.

Security evaluation laboratories assess the resistance of cryptographic implementations against such attacks. However, this process is costly and must strike a balance between thoroughness and practical limitations on time, budget, data, and computational resources. Deep learning-based side-channel analysis (DL-SCA) is attractive in this context because neural networks can learn leakage characteristics directly from traces, reducing the need for manual preprocessing and explicit statistical assumptions. At the same time, deep learning introduces new costs, caused by sensitivity to neural network hyperparameter selection, instability, and overfitting in its training process.

The central problem addressed in this thesis is the tension between the benefits and costs of deep learning in side-channel evaluation. On the one hand, deep learning can reduce evaluation effort by relaxing assumptions about leakage models and reducing dependence on known data. On the other hand, it can make evaluation more expensive due to model selection, hyperparameter tuning, and the risk of overfitting. This thesis investigates how DL-SCA can be made more practical, reliable, and cost-effective for security evaluation workflows.

To this end, the thesis studies several strategies for improving DL-SCA without relying on excessive hyperparameter tuning. It examines the impact of increasing the amount of training data, regularization techniques, and ensemble learning. These approaches aim to improve generalization, robustness, and attack stability under realistic evaluation constraints. The thesis also investigates two deep learning approaches that relax major assumptions in classical SCA: leakage model-flexible DL-SCA, which avoids relying on fixed leakage models such as Hamming weight or identity, and deep learning-based blind SCA, which reduces dependence on plaintext or ciphertext by learning from noisy labels.