How To Find ML Bugs That Expose Data and Bias Outcomes

Abstract:
Systems and services that rely on data to provide functionality are widespread, as a growing number of high-profile success stories drives their adoption into new domains. Increasingly, the technology underpinning this trend is deep learning, which has enabled new applications that had previously eluded traditional software development methods. However, this development has also been met with concerns around the privacy of individuals’ data, and the potential for these systems to discriminate in unintended and harmful ways. In this talk, I will show how privacy and fairness in ML applications concerns can be related through the lens of protected information use, and show that tools developed to help characterize ML models’ use of such information can uncover new types of “bugs” that expose private training data and lead to unwarranted discrimination. Finally, I will discuss promising techniques that address these issues through novel data representations and model post-processing, leading to ML applications that solve important problems without jeopardizing the privacy and fairness concerns of their users.
Biography:
Matt Fredrikson is an Assistant Professor of Computer Science at Carnegie Mellon University, where he joined in 2015 after receiving his PhD from the University of Wisconsin, Madison. His research aims to make ML-based systems more transparent and reliable by bringing rigorous techniques to bear on problems of fairness, privacy, and security. He is the recipient of a NSF CAREER award, and has received multiple best paper awards for his work on private machine learning.