Human-centric machine learning: on the preservation of individual privacy and fairness
日期：2020/07/29 - 2020/07/29
学术讲座: Human-centric machine learning: on the preservation of individual privacy and fairness
主讲人: Xueru Zhang, University of Michigan
地点：via Zoom (Meeting ID: 63713973191, Password: 5488)
Machine learning (ML) techniques have seen significant advances over the last decades and are playing an increasingly critical role in people's lives. While their potential societal benefits are enormous, they can also inflict great harm if not developed or used with care. In this talk, I will focus on two critical ethical issues in ML systems: discrimination and privacy violation, and present mitigating solutions in various scenarios.
On the fairness front, ML models that are trained with data from multiple demographic groups can inherit pre-existing biases in the data such that the models exhibit discrimination against already-disadvantaged or marginalized social groups. In the first part of my talk, I will present my work on evaluating the long-term impacts of ML (fair) decisions on different social groups in a sequential framework. Specifically, I will illustrate how the biases can perpetuate and worsen over time in ML systems, and how imposing common fairness criteria that intend to promote fairness may still lead to undesirable pernicious long-term effects. I will then discuss an approach to remedying this issue.
On the privacy front, when ML models are trained over individuals’ personal data, it is critical to preserve each individual’s privacy while maintaining a sufficient level of model accuracy. In the second part of the talk, I will illustrate two key ideas that can be used to balance an algorithm’s privacy-accuracy tradeoff: (1) reuse intermediate computational results to reduce information leakage; and (2) improve algorithmic robustness to accommodate more randomness. I will present a randomized, privacy-preserving algorithm that leverages these ideas in the context of distributed learning, where multiple entities collaboratively work toward a common learning objective through an interactive and iterative process of local computation and message passing. It is shown that our algorithm’s privacy-accuracy tradeoff can be improved significantly over existing algorithms.
Xueru Zhang is a Ph.D. candidate in the Department of Electrical and Computer Engineering at the University of Michigan, advised by Mingyan Liu. She received her M.Sc. degree in Electrical and Computer Engineering from the University of Michigan in 2016 and B.Eng. degree in Electronic and Information Engineering from Beihang University (BUAA), Beijing, China, in 2015. Her research lies at the intersection of machine learning, optimization, statistics and economics, including topics such as data privacy, algorithmic fairness and security economics. She was a recipient of Rackham Predoctoral Fellowship in 2020.