Advances in artificial intelligence (AI) come with the promise of life-changing improvements in technologies such as autonomous vehicles, facial recognition, and health diagnostics. However, the trustworthiness of AI has become a major problem and has restricted its applications. In this talk, I will introduce my works on promoting AI trustworthiness from the perspective of improving AI robustness and preserving data privacy in the AI workflow.

In the first part of the talk, I will present my work on Trojan model detection in the data-free regime. Trojan attack, a training phase threat, maliciously tempers training data to affect neural network training and can significantly harm real-life AI systems in downstream applications. Here, I will introduce a novel nonconvex optimization-based detector that leverages the activation of hidden neurons and is built through the reverse engineering of Trojan triggers in the input-agnostic cases. The detector can efficiently reveal Trojan triggers, detect Trojan models, and even find target Trojan classes. In the second part of the talk, I will present a quantization approach associated with a provable recovery method to preserve privacy in the data collection/processing phase without hindering data accuracy in model learning. The approach guarantees that any intruder accessing a small amount of data cannot reveal accurate information even with the knowledge of the quantization rule. Along the way, I will introduce a distributed algorithm to enhance security and demonstrate how to improve data accuracy with a multi-copy strategy and a tensor method. The talk will also include a brief introduction to other branches in my trustworthy AI research and my research areas beyond AI trustworthiness.

Dr. Ren Wang is a postdoctoral research fellow and a lecturer in the Department of Electrical Engineering and Computer Science at the University of Michigan. He obtained his Ph.D. in 2020 from the Department of Electrical, Computer, and Systems Engineering at Rensselaer Polytechnic Institute, where he was an IBM AI Horizons Fellowship recipient and received the Rensselaer’s Founders Award of Excellence. Before that, he received his bachelor’s degree and master’s degree in Electrical Engineering from Tsinghua University in 2013 and 2016, respectively. He also has industry internship experience at IBM Research AI and ABB. His research interests include Trustworthy Machine Learning, High-Dimensional Data Analysis, Bio-Inspired Machine Learning, and Robustness/Optimization on Smart Grids.