Publication Type
PhD Dissertation
Version
publishedVersion
Publication Date
1-2026
Abstract
Machine learning plays a critical role in solving real-world problems. Yet, these remarkable achievements may often be overshadowed by safety issues, such as the lack of robustness or fairness. These problems pose safety and ethics threats to many critical systems built on machine learning. Numerous approaches have been proposed to address these issues by empirically suppressing failures. However, these attempts tend to fail under adaptive attacks, thus calling for methods with formal guarantees. On the other hand, existing methods with guarantees typically suffer from noticeable utility drop. The exact mechanism between these safety demands and accuracy loss remains unclear, which also hinders the development of reasonable remedies. In this dissertation, the aim is to study problems associated with certified AI Safety, with a particular focus on robustness and fairness.
In the first work, we adopt the Bayes error for robustness analysis. We investigate the limit of certified robust accuracy, taking into account data distribution uncertainties. We establish an upper bound for certified robust accuracy, considering the distribution of individual classes and their boundaries. Our theoretical results are empirically evaluated and are consistent with the limited success of existing certified training results. Our second work continues the idea and extends it to probabilistic robustness. We find that while Bayes uncertainty does affect probabilistic robustness, its impact is smaller than that on deterministic robustness. This reduced Bayes uncertainty allows a higher upper bound on probabilistic robust accuracy than that on deterministic robust accuracy. In our third work, we propose T&T, an approach to achieving high accuracy with certified probabilistic robustness. T&T has two parts which together achieve our goal, i.e., a probabilistic robust training with an additional goal of minimising divergence variance in a given vicinity and a runtime inference via robustness testing. In our fourth work, we focus on individual fairness. Individual fairness is another important constraint for systems, and it can be derived in the form of robustness. This work proposes a correct-by-construction training approach that formally guarantees perfect fairness throughout model development, i.e., utility improvement. Together, these works offer a comprehensive framework for understanding certified AI.
Degree Awarded
PhD in Computer Science
Discipline
Artificial Intelligence and Robotics | Computer Sciences
Supervisor(s)
SUN, Jun
First Page
1
Last Page
137
Publisher
Singapore Management University
City or Country
Singapore
Citation
ZHANG, Ruihan.
Certifying AI with robustness and fairness. (2026). 1-137.
Available at: https://ink.library.smu.edu.sg/etd_coll/829
Copyright Owner and License
Author
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.