学术交流

Optimizing Distributed Machine Learning System: Towards Better Security, Utility, and Efficiency

发布者：梁慧丽发布时间：2025-11-28浏览次数：10

报告人

Di Chai

The Hong Kong University of Science and Technology

时间

2025年6月17日 星期二

下午 14:00-15:00

地点

602 会议室


Abstract


Distributed machine learning systems leverage distributed data and computational resources to support applications across finance, healthcare, and smart cities, with core objectives encompassing data privacy, model utility, and infrastructure efficiency. Our work makes three key contributions: (1) For security in distributed matrix factorization (MF) - a critical yet underexplored area compared to deep learning - we identify security vulnerabilities and develop a lightweight protection framework combining homomorphic encryption (HE) and mutual information minimization, protecting raw data and reducing inference attack accuracy by 16%; (2) For utility in traffic prediction, we pioneer the use of graph neural networks (GNNs) to model complex urban data and merge area correlations through multi-graph fusion, achieving a 25% reduction in prediction error; (3) For system efficiency, our algorithm-system co-design enables billion-scale data processing through domain-specific matrix protection, CPU kernel optimizations, and communication complexity analysis, delivering >23,000× throughput gains over conventional HE with lossless accuracy. Building on these advances, future work will focus on developing secure, accurate, and efficient LLM systems from a data-centric perspective.


Biography


Di Chai received his Ph.D. in Computer Science and Engineering from the Hong Kong University of Science and Technology (HKUST) in March 2025, under the supervision of Prof. Chen Kai and Prof. Yang Qiang. His research focuses on building secure, accurate, and high-performance machine learning systems. His key contributions include: (1) developing a novel secure matrix factorization framework for privacy-preserving recommender systems and genomic data analysis; (2) pioneering multi-graph fusion GCN techniques in urban computing to model dynamic and complex traffic flow patterns; and (3) designing an efficient system spanning optimizations in computation, communication, and storage to enable efficient billion-scale data processing and decentralized computing. During his Ph.D., he published 13 high-quality papers (including 6 CCF-A papers) in top-tier venues such as KDD, ATC, SIGMOD, TKDE, and IEEE S&P, with over 1,000 citations. Moving forward, he plans to focus on data-centric large language model (LLM) systems by improving training data quality and optimizing contextual management during inference to support secure, accurate, and efficient large-scale AI models.

上海财经大学 计算经济交叉科学教育部重点实验室©版权所有

搜索
您想要找的