COMMITTEE CHAIR: Dr. Annamalai Annamalai

TITLE: DEEP LEARNING-BASED SIDE-CHANNEL EVALUATION ACROSS CLASSICAL AND POST-QUANTUM CRYPTOGRAPHIC IMPLEMENTATIONS

ABSTRACT: A cryptographic algorithm that is mathematically secure can still be broken since its hardware implementation leaks secret information through power consumption or electromagnetic emission. These vulnerabilities are applicable to widely deployed classical encryption schemes as well as emerging post-quantum schemes which are standardized to provide quantum resistance. This thesis evaluates how deep learning changes such side-channel attacks across both through a unified study of four attacks, spanning a classical block cipher and two protected post-quantum schemes, and tests learned methods against the conditions that separate a laboratory result from a practical attack. Profiled deep-learning attacks recover keys efficiently under favorable laboratory conditions. However, their behavior is far less well understood when traces are misaligned, when the attacked key differs from the one used in training, and when the leakage is faint and measured on real hardware. The thesis asks whether learned attacks remain effective under these conditions, and where they reach their limits. The four attacks were evaluated on physical devices, targeting AES-128 on a microcontroller and the post-quantum schemes ML-KEM and HQC on field-programmable gate arrays. Each attack was measured by guessing entropy and the number of traces needed to recover a key. On the classical cipher, an ensemble of convolutional networks stabilized recovery against a timing countermeasure. Selecting the most informative trace samples then produced a far smaller model that recovered keys it had never seen, which establishes that the exploited leakage is a property of the device rather than of any single key. On the post-quantum targets, the limiting factor was not the sophistication of the model but the distance between training and deployment. The smallest network often outperformed an attention-based transformer many times its size, and the gap between simulated and real measurements bounded every learned attack. On a protected implementation whose leak defeated standard attacks, knowledge distillation nonetheless recovered the targeted secret byte. Across the four studies, what limited a practical attack was not the complexity of the model but the distance between training and deployment conditions. Deep learning helped most when a method was aimed at a specific obstacle, and the least when complexity was added for its own sake.

Keywords: Post-quantum cryptography, power and side-channel analysis, advanced encryption standard, deep learning, electromagnetic analysis, Hamming quasi-cyclic, knowledge distillation, module-lattice-based key-encapsulation mechanism

Room Location: Electrical Engineering Conference Room 315D