Understanding and Improving the Trustworthiness of Machine Learning

March 11, 2025
Categories: Events, Seminar Series

Zitao Chen
Electrical and Computer Engineering (ECE)
University of British Columbia (UBC)

March 19 2025
11-12 WWH 335

Machine Learning (ML) has seen increasing use in many high-stakes scenarios across our society. Despite their impressive performance in typical operations, ML models are subject to catastrophic failures, such as information leakage or safety violations. This talk will examine three prominent issues in building trustworthy ML systems: privacy, accountability, and safety. It is well known that ML models are prone to leaking sensitive information. But to
what extent can such leakage be deliberately exacerbated by malicious parties? To answer this, I will describe how ML models can be manipulated to leak substantial information in unconventional ways that are hard to detect, even with state-of-the-art privacy techniques. Even without being proactively manipulated, ML models can still leak considerable information, and I will discuss how we can mitigate this. To continue, I will briefly describe how malicious privacy attacks can be reimagined for ordinary users such as personal artists to detect the unauthorized use of their data in ML training. Once deployed into operation, ML models also face potential safety risks. To this end, I will discuss how we can enable the safe operation of ML even under the presence of hardware faults. Finally, this talk will conclude by outlining future directions for advancing the trustworthy development and deployment of ML.

Zitao Chen’s research focuses on analyzing the unexpected failure modes in machine learning and developing solutions to build trustworthy machine learning systems. He is a German DAAD AI-Net Fellow, and a UBC Public Scholar. He has won a best paper award and best paper award runner-up in DSN’21. His work also led to industrial adoption and was selected for IEEE Top Picks in Test and Reliability (2024) as one the most impactful publications in the area of Computer Systems Reliability from 2018 to 2024.