More about HKUST
Secure and Practical Federated Matrix Factorization
The Hong Kong University of Science and Technology Department of Computer Science and Engineering PhD Thesis Defence Title: "Secure and Practical Federated Matrix Factorization" By Mr. Di CHAI Abstract: Matrix factorization (MF) is an essential primitive to support various applications including recommender systems, genetic studies, topic modeling, financial applications, etc. While most MF-based applications deal with large-scale sensitive data, e.g., users' shopping history in the recommender system and genome data in genetic studies, they encounter the isolated-data-islands problem. Restricted by the privacy-preserving regulations, it is challenging to share the data across different parties. Consequently, the data are in the form of isolated islands and conventional MF that requires data collected centrally cannot work. To solve this problem, federated MF has been proposed to decompose data distributed among different parties. While pioneered studies have shown the feasibility of federated MF, they are limited in terms of privacy, utility, or efficiency. In this thesis, we first define the problem of federated MF and its two subproblems: federated sparse MF and federated dense MF. We then present the targets achieved by existing works and propose three new research questions. In the first question, we focus on the MF-based recommender system that deals with sparse matrices. Existing works exchange gradients in plaintext, raising significant privacy concerns. Our first contribution is mathematically proving that gradients leak raw data, leading us to protect the gradients using homomorphic encryption. However, even with the gradient values secured, the indexes of the gradients can still leak information (e.g., used for inference attacks) due to the sparsity of the rating matrices. To address this, we propose an obfuscation-based method to defend against inference attacks with bounded utility loss and efficiency overhead. In the second question, we focus on federated dense MF which supports applications such as genetic studies, topic modeling, and financial applications. These applications handle large-scale dense matrices and require high accuracy. Existing works either suffer from utility penalties because of leveraging differential privacy or face severe efficiency issues due to leveraging homomorphic encryption. We propose a practical lossless federated singular vector decomposition (SVD) system that is capable of decomposing billion-scale data. Specifically, we propose lossless masking-based protection tailored for federated SVD and improve the efficiency through extensive system optimizations. The evaluation results demonstrate that our proposed solutions outperform existing studies and significantly improve the practicality of MF-based real-world applications. In the third question, we concentrate on decentralized federated MF. Most of the existing solutions rely on third-party servers, which compromises the system's security. Although some works have tried to remove the external servers using HE, they suffer from severe efficiency issues. We remove the external servers using a novel lightweight matrix protection and achieve high efficiency through vest computational and communication optimizations. Date: Friday, 17 January 2025 Time: 12:30pm - 2:30pm Venue: Room 3494 Lifts 25/26 Chairman: Dr. Shiheng WANG (ACCT) Committee Members: Prof. Qiang YANG (Supervisor) Prof. Kai CHEN (Supervisor) Dr. Qifeng CHEN Prof. Ke YI Dr. Can YANG (MATH) Prof. Chuan WU (HKU)