Towards Effective Countermeasures Against Deep Learning Side-Channel Attacks
Time: Thu 2026-12-10 13.00
Location: Hörsal, Malvinas väg 10
Video link: https://kth-se.zoom.us/s/64551861470
Language: English
Subject area: Information and Communication Technology
Doctoral student: Martin Brisfors , Elektronik och inbyggda system
Opponent: Associate Professor Francesco Regazzoni, University of Amsterdam
Supervisor: Elena Dubrova, Elektronik och inbyggda system
QC 20261117
Abstract
The emergence of deep learning-based side-channel attacks has revealed that many existing cryptographic countermeasures have vulnerabilities that older statistical analysis could not exploit. This undermines established security assumptions and highlights a need to re-evaluate current countermeasure designs. Ideally, new countermeasures would be developed to combat the emerging threat of deep learning attacks.
To address these vulnerabilities, this thesis evaluates novel attack methodologies against protected implementations. It presents an attack that combines oversampling and synchronization, demonstrating that increasing the oversampling factor from 8 to 20 reduces the protection of the proposed asynchronous clock-randomization countermeasure by over 99.99%. It also identifies a vulnerability in the initial Clock Randomization with Core Duplication design shown in this thesis. It proves that concurrent clock rising edges create leakage that enables a successful attack. It also presents a secret key recovery attack on second-order masked Module Lattice Key Encapsulation Mechanism, identifying leakage arising from conversions between Boolean and arithmetic masking, and introduces an optimal ciphertext selection process to maximize attack efficiency.
In addition, this thesis presents countermeasure designs developed to withstand deep learning-based attacks. The thesis proposes an asynchronous clock-randomization design intended to maximize trace misalignment. To address the vulnerabilities identified in the clock-randomization design, the thesis also introduces the Clock Randomization with Core Duplication countermeasure. This hides leakage by running a duplicate algorithm using a dummy encryption key, which prevents synchronization attacks. The thesis further shows and adaptation of this architecture to secure a Module Lattice Key Encapsulation Mechanism implementation using synchronous clock switching to maintain system stability. Finally, the thesis shows an updated design of the countermeasure that incorporates an additional mixed-mode clock manager. This prevents the common-mode vulnerability shown in the paper. The new design also features frequency capping which prevents the instability present in the original design.
This thesis demonstrates that Clock Randomization with Core Duplication resists deep learning-based attacks while maintaining a more favorable overhead than higher-order masking. It also highlights the risk of overestimating the security of lightweight, pure clock-randomization schemes. Together, these findings emphasize the necessity of rigorous empirical validation and offer new directions for future countermeasure designs.