Towards Robust Machine Learning under Distribution Shift and Adversarial Attack

October 27, 2021 (Xintao Wu)

As big data and AI technologies are deployed to make critical decisions that potentially affect individuals (e.g., employment, college admissions, credit, and health insurance), there are increasing concerns from the public on privacy, fairness, safety, and robustness issues of data analytics, collection, sharing and decision making. In this talk, we first overview our social awareness research, in particular, on how to mitigate side effect of enforcing one social concern on another, and how to address multiple social concerns simultaneously. We then focus on robustness of machine learning under two representative scenarios, distribution shift and adversarial attack. In the former scenario, we present robust learning based on kernel reweighing and Heckman model. In the second scenario, we present adaptive defense that purposely leverages multiple types of adversarial samples to learn the context information in the training. We conclude the talk with some future research directions.