Bo Li (UIUC) - Certifiably Robust Learning via Knowledge-Enabled Logical Reasoning

As intelligent systems become pervasive, safeguarding their trustworthiness is critical. Standard machine learning (ML) approaches often assume that the training and test data follow similar distributions, without taking into account the possibility of adversaries manipulating either distribution or natural distribution shifts. This can lead to severe trustworthiness issues in ML. For instance, our past work has shown that motivated adversaries can circumvent anomaly detection or other machine learning models at test-time through evasion attacks, or inject well-crafted malicious instances into training data to induce errors through poisoning attacks. In this talk, I will provide a brief overview of my research on trustworthy machine learning, including robustness, privacy, generalization, and their underlying interconnections, with a focus on robustness. In particular, I will first discuss the current state of the art in certifiably robust defenses based on purely data-driven learning approaches and their limitations. I will then present our certifiably robust learning via knowledge-enabled logical reasoning approach, including a thorough analysis of its properties. I will demonstrate that it is possible to 1) certify the robustness of this end-to-end approach, which significantly improves the SOTA certified robustness, 2) prove this approach is more robust than a single ML model under mild conditions, 3) make it scalable for a variety of downstream tasks, and this approach is agnostic to the tasks, data, and attacks being used.