Improving the Efficiency of Federated Recommender System

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Improving the Efficiency of Federated Recommender System"

By

Mr. Liu YANG


Abstract:

Recommender system (RecSys) provides personalized recommendations services
across various applications such as e-commerce, online video, and social
networking, making considerable profits for companies. The technologies of
RecSys have evolved from shallow to deep and large models, highlighting the
transition towards more complex models to better capture user-item
interactions. However, the centralized nature of traditional RecSys faces
challenges in data privacy.

Federated recommender systems (FedRec) have emerged as a promising paradigm to
address the data silo and privacy issues by enabling collaborative model
training and inference without sharing raw data. However, when combined with
security-enhancing techniques such as secret sharing or homomorphic
encryption, FedRec suffers from severe efficiency challenges in terms of
computation, communication, and convergence.

This proposal therefore focuses on improving the efficiency of FedRec,
systematically studying three representative scenarios: (1) horizontal
FedRec in cross-device settings dominated by shallow models, (2) vertical
FedRec in cross-silo settings with deep models, and (3) transfer FedRec in
client–server settings with large models.

To this end, we develop a series of methods tailored to each scenario. For
horizontal FedRec, we propose FedMMF, which leverages personalized masks and
adaptive aggregation to reduce the quadratic complexity of
secret-sharing–based protocols while maintaining accuracy. For vertical
FedRec, we design PackVFL, an efficient packed homomorphic encryption
framework that significantly accelerates secure matrix multiplication and
supports practical deployment with deep models. For transfer FedRec, we
introduce FedShuffle, a lightweight mechanism to protect user queries to
large remote models, addressing efficiency bottlenecks of existing
cryptographic and textual obfuscation methods.

By targeting efficiency as the central problem, this proposal highlights a
unified perspective across different categories of FedRec, bridging shallow,
deep, and large models. The proposed methods advance the practicality of
FedRec and pave the way for scalable, efficient, and privacy-preserving
recommender systems in real-world applications.


Date:                   Friday, 31 October 2025

Time:                   10:00am - 12:00noon

Venue:                  Room 3494
                        Lifts 25/26

Chairman:               Prof. Ling SHI (ECE)

Committee Members:      Prof. Kai CHEN (Supervisor)
                        Prof. Qiang YANG (Co-supervisor)
                        Prof. Qiong LUO
                        Dr. Binhang YUAN
                        Prof. Can YANG (MATH)
                        Dr. Yang LIU (PolyU)