Confronting the Privacy Leak Epidemic in the Machine Learning Era

ABSTRACT

As machine learning becomes increasingly integrated into our daily lives, concerns about data privacy behind these intelligence services have grown significantly. Advanced ML models often rely on sensitive user information to deliver high-performance, personalized services. This growing reliance makes private user data an attractive target for exploitation, leading to various forms of privacy leakage arising from interactions with ML applications. This presentation will delve into two critical dimensions of these data privacy threats. First, we explore how massive amounts of private user data are collected to fuel advanced ML services. We uncover that mobile notification services—used daily for receiving ads, news, and alerts—have become a significant vector for sensitive data leakage. By dissecting the mobile notification ecosystem, we analyze its implementation process and expose real-world illicit data collection practices, which impact billions of users globally. Second, we analyze vulnerabilities within ML models that can further compromise user privacy. Specifically, we examine membership inference attacks, highlighting how ML models can leak their training datasets containing sensitive data. Furthermore, we propose a novel defense mechanism to counteract link-stealing in Graph Neural Networks, a cutting-edge model powering social media platforms. Through this analysis, our efforts not only inspire further in-depth research into privacy issues within the ML services ecosystem but also raise public awareness about the importance of safeguarding privacy in the era of machine learning.

BIOGRAPHY

Jiadong Lou is a final year Ph.D. candidate in the Department of Computer & Information Sciences at the University of Delaware, advised by Dr. Xu Yuan. His research interests include cybersecurity, machine learning, AI for science, and cyber-physical systems. During his Ph.D. studies, he has produced 20 outcomes, with 15 published in prestigious venues such as IEEE S&P, USENIX Security, ACM CCS, and IEEE INFOCOM. Jiadong Lou also received the Best Paper Award and Distinguished Paper Award at DSN 2023, as well as the IEEE CNS 2024 Best Paper Runner-up  Award.