Safeguarding Privacy in Machine Learning: Challenges and Innovations from Edge to Cloud

Date: 2024/08/01 - 2024/08/01

Academic Seminar: Safeguarding Privacy in Machine Learning: Challenges and Innovations from Edge to Cloud

Speaker: Zhifeng Jiang, Hong Kong University of Science and Technology

Time: 09:00-10:00, Room 454, August 1, 2024 (Beijing Time)

Location: Room 454, JI Longbin Building

Abstract

The abundant data and computing resources on client devices, such as smartphones and AI cameras, have made machine learning (ML) on the edge not only possible but increasingly prevalent. However, ensuring that model building leverages distributed data without compromising users' privacy remains a significant challenge. In this talk, I will present two works addressing privacy vulnerabilities in the current workflow of edge learning. First, I will discuss how privacy-preserving ML models are typically built under the coordination of a server and the collaboration of participating clients. I will introduce a new design that maintains target privacy even in the face of passive failures, such as client dropout. Next, I will explore scenarios involving active attackers, where privacy can be breached through server-client collusion, and present a novel design that prevents the server from exploiting such collusion. Finally, given that training and deploying large, productive models like GPT-4 or DALL·E 2 on edge devices is often infeasible due to substantial resource demands and business constraints, I will conclude with an explorative discussion on private learning in the cloud for a broader vision.

Biography

Zhifeng Jiang is a recent Ph.D. graduate from the Department of Computer Science and Engineering at the Hong Kong University of Science and Technology, where he worked under the guidance of Prof. Wei Wang and Prof. Bo Li. He earned his B.Eng. degree from the Department of Computer Science and Technology at Zhejiang University in 2019. Zhifeng's research interests span the broad area of machine learning systems, with a particular focus on ensuring privacy, security, and trustworthiness. His thesis research centered on enhancing user privacy and training efficiency for edge learning, while his ongoing research also expanded to the privacy and security aspects of machine learning in the cloud, especially large generative models. He was a co-recipient of the Best Paper Award Runner-Up at the IEEE International Conference on Distributed Computing Systems (ICDCS).