PhD Dissertation Defense by ZHANG Ruihan | Certifying AI with Robustness and Fairness

Certifying AI with Robustness and Fairness

ABOUT THE TALK

Machine learning is widely used in real-world applications, but its deployment raises serious safety concerns, including limited robustness and fairness. These shortcomings pose significant ethical and practical risks, especially in safety-critical systems. Although many empirical techniques aim to address these issues, they often fail under adaptive attacks, underscoring the need for methods with formal guarantees. Existing certified approaches, however, typically suffer from substantial utility loss. Moreover, the mechanisms underlying the trade-off between safety guarantees and accuracy are not well understood, limiting progress in this area. This dissertation investigates certified AI safety, with a focus on robustness and fairness.

The first work leverages Bayes error to analyze robustness, studying the fundamental limits of certified robust accuracy under data distribution uncertainty. It derives an upper bound based on class-conditional distributions and their decision boundaries, with empirical results validating the theory and explaining the limited effectiveness of current certified training methods. The second work extends this framework to probabilistic robustness, demonstrating that while Bayes uncertainty still plays a role, its effect is weaker than in deterministic robustness, resulting in a higher achievable upper bound. Building on these findings, the third work introduces T&T, a method that achieves high accuracy with certified probabilistic robustness by combining probabilistic robust training with a runtime robustness testing procedure. The fourth work addresses individual fairness, framing it as a robustness problem and proposing a correct-by-construction training approach that guarantees perfect individual fairness while enabling improved utility. Collectively, these contributions offer a principled framework for understanding limits, trade-offs, and opportunities in certified AI safety.

SPEAKER BIOGRAPHY

Ruihan ZHANG is completing her Ph.D. in Computer Science at SMU under the supervision of Prof. Sun Jun. She previously earned her Bachelor of (with Honours) from the Singapore University of Technology and Design. Her research focuses on trustworthy AI, where she uses formal methods to study and verify key properties such as robustness, fairness, and reliability in neural networks. Her work has been published in leading venues including CAV, AAAI, OOPSLA, and TReliab. Ruihan also gained industry experience through the VISA Global Intern Program and has been recognised with awards such as the SMU Research Excellence Awards and scholarships from MOE and SUTD.