2024년 11월 1일 10시에 진행되는 중앙대학교 AI대학원 CAU-AI핵심 기술세미나 진행 안내드립니다.

이번 세미나는 ‘야밤의 공대생 만화’의 작가이자 현재 펜실베니아 주립대 교수이신 맹기완 교수님을 초청하여 진행됩니다.

관심있는 분들의 많은 참여 부탁드립니다.

Presenter : 맹기완 조교수 (Penn State Univ)

- Postdoc Researcher. Facebook AI Research (FAIR; 2021 -- 2022)

- Research Intern. Facebook AI Research (FAIR; 2020.5 -- 2020.12)

- Research Intern. Microsoft Research (2019.5 -- 2019.8)

- Kiwan Maeng is a Charles K. Etner Early Career assistant professor at Pennsylvania State University,

  computer science and engineering department. His main research interests lie in co-optimizing ML

  algorithms and systems for improved privacy, accuracy, and efficiency. Prior to joining Pennsylvania State University, he worked at Meta AI Research (formerly FAIR) SysML team.

Date & Time : Nov. 01, 2024. 10:00 (KST)

Place : Online Zoom ( https://cau.zoom.us/j/82608427445 )

Title : Privacy-Preserving Machine Learning Across the User Device and the Cloud

 Many modern machine learning (ML) systems operate across users’ edge devices and the cloud server. During training/inference, sensitive user data is revealed to the cloud, posing potential privacy threats. While many privacy-preserving training and inference algorithms have been proposed to protect sensitive users data from being misused, the field is still nascent. Many studies have concentrated solely on the algorithm or the system, neglecting potential benefits or issues when considering cross-layer optimization and integration.

In this talk, I will focus on two recent projects that try to enable privacy-preserving ML across the edge-cloud with algorithm-system codesign. The first work enables private training/inference using instance encoding—a technique that encodes data to enhance privacy while preserving utility—with theoretically-meaningful privacy for the first time to the best of our knowledge. We propose a theoretical framework that can lower bound the error when an arbitrary attacker tries to invert the encoding using diagonal Fisher information leakage (dFIL), and show how one can build a collaborative training/inference system using the framework. The second work enables private ML inference using multi-party computation (MPC). We show how the cryptographically-safe but slow MPC algorithm can be optimized while adding only minimal error, through an approximated ReLU estimation on a reduced integer ring.

*This seminar was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (2021-0-01341, Artificial Intelligence Graduate School Program(Chung-Ang University))