PhD Dissertation Proposal by ZHANG Ruihan | Accuracy Upper Bounds in Constrained Machine Learning

ABOUT THE TALK

More than a decade has passed since the adversarial example was discovered. However, the robustness issue remains a major challenge for traditional neural networks as well as large models (including large language models and multi-modal models). In this talk, I will share a series of research work based on formal methods to address this issue. From a theoretical perspective, we explore the aim of perfect robustness (i.e., the absence of adversarial examples) is unreachable and further investigate whether probabilistic robustness (i.e., the limited presence of adversarial examples) can be a practical compromise. In practice, we propose a method that combines lightweight neural training with inference to achieve theoretically guaranteed probabilistic robustness for the network.

ABOUT THE SPEAKER

Ruihan Zhang is a PhD Candidate in Computer Science at the SMU School of Computing and Information Systems, supervised by Prof. Jun SUN.  Her research is focused on machine learning robustness and fairness.