Join Assistant Professor Minhao Cheng for an IST Research Talk titled "Post-Hoc Security in Machine Learning Systems." He'll discuss his current research projects, new findings, and opportunities for collaboration. All talks are free and open to the public.

Abstract 

Machine learning systems, while powerful, remain vulnerable to diverse adversarial threats. Existing defenses, even those widely adopted, are frequently bypassed by more sophisticated attacks. In this talk, I will first demonstrate the vulnerability of a common defense technique against backdoors. Specifically, I will show how machine learning systems can swiftly relearn malicious behavior through minimal exposure, even from simple queries. To bolster machine learning security against such advanced threats, I will introduce my research on building digital forensics frameworks. These frameworks offer post-breach protection, complementing traditional defenses. They enable the tracing of AI-generated content origins, assisting regulators in ensuring the safe and responsible use of machine learning. Our works provides a new perspective, suggesting a shift from a purely preventive security mindset toward a more comprehensive approach that includes robust post-breach analysis and response capabilities.

About the Speaker

Minhao Cheng is an assistant professor in the Penn State College of Information Sciences and Technology. He works at the intersection of security and machine learning, with a particular focus on automated, efficient, and trustworthy machine learning systems. Before joining Penn State, Cheng was an assistant professor of computer science and engineering at Hong Kong University of Science and Technology. He holds a doctoral degree in computer science from the University of California, Los Angeles.