Practical Attacks and Defenses in Deep Learning Systems

Abstract: Despite the fact that deep learning systems have been used for a wide range of applications, such as power systems, computer vision, and healthcare, the lack of robustness has resulted in ever-growing concern over how people can trust these models in applications that require high security.  In this talk, I will begin by discussing Trojan attacks, also known as training-phase backdoor attacks, which pose a significant threat to machine learning systems. These attacks strategically embed backdoor patterns into a well-trained model, enabling attackers to manipulate the machine decision-making process.  I will subsequently explore practical defenses against these Trojan attacks, which include techniques for detecting backdoored models and data, and for repairing them.  Towards the end of the talk, the focus will shift to inference-phase adversarial attacks.  These are subtle yet effective techniques that manipulate test data to mislead the decisions of machine learning systems.  Following this, I will delve into recent methods developed to enhance models' intrinsic robustness against these adversarial attacks, leveraging concepts such as robust mode connectivity.  The talk will also highlight the connections between the introduced attacks and defenses, with power systems and optimization.

Speaker’s Bio: Dr. Ren Wang is an Assistant Professor in the Department of Electrical and Computer Engineering at Illinois Tech.  Before joining Illinois Tech, he was a postdoctoral research fellow in the Department of Electrical Engineering and Computer Science at the University of Michigan.  His research interests include trustworthy machine learning, high-dimensional data analysis, and smart grids. He is also the recipient of the 2023 Oak Ridge Associated Universities Ralph E. Powe Junior Faculty Enhancement Award.