Secure Federated Graph-Filtering for Recommender Systems

TL;DR

The proposed decentralized frameworks achieve comparable accuracy to centralized state-of-the-art systems while ensuring data confidentiality and maintaining low communication costs, and highlight the potential for privacy-preserving decentralized architectures to bridge the gap between utility and user data protection in modern recommender systems.

Abstract

Recommender systems often rely on graph-based filters, such as normalized item-item adjacency matrices and low-pass filters. While effective, the centralized computation of these components raises concerns about privacy, security, and the ethical use of user data. This work proposes two decentralized frameworks for securely computing these critical graph components without centralizing sensitive information. The first approach leverages lightweight Multi-Party Computation and distributed singular vector computations to privately compute key graph filters. The second extends this framework by incorporating low-rank approximations, enabling a trade-off between communication efficiency and predictive performance. Empirical evaluations on benchmark datasets demonstrate that the proposed methods achieve comparable accuracy to centralized state-of-the-art systems while ensuring data confidentiality and maintaining low communication costs. Our results highlight the potential for privacy-preserving decentralized architectures to bridge the gap between utility and user data protection in modern recommender systems.

Figures (5)

• Figure 1: High-level system model for PriviRec. Each user $u$ holds local interaction data $\bm{R}^{(u)}$. A central server coordinates the overall protocol. It uses Secure Aggregation to receive masked (confidential) partial sums from the clients and homomorphically sums them to compute unmasked global filters (e.g. $\tilde{\bm{P}}, \bm{F}_{IDL}$). It then broadcasts the aggregated results back to the clients.
• Figure 2: Sequence diagram for the Distributed Secure Power Method algorithm. The server and clients collaboratively compute the eigen-space of $\tilde{\bm{P}}$.
• Figure 3: NDCG of GF-CF on Gowalla and Yelp2018, when its components are either computed using PriviRec, or PriviRec-$k$. We vary $k$ between 256 and 3584 in increments of 256.
• Figure 4: Sequence diagram for the PriviRec Algorithm. The algorithm involves secure aggregation, server and clients-side computations to return the approximate item-item matrix $\tilde{\bm{P}}$ and the ideal low-pass filter $\bm{F}_{IDL}$.
• Figure 5: Sequence diagram for the PriviRec-$k$ Algorithm. The algorithm involves collaborative computation, server and client-side computation to return the approximate item-item matrix $\tilde{\bm{P}}_k$ and the ideal low-pass filter $\bm{F}_{IDL}$.

Theorems & Definitions (8)

• Theorem 3.1
• proof
• Theorem 3.2
• proof : Derivation
• Lemma 3.4
• proof : Proof of Lemma \ref{['secanallemma']}.
• proof : Communication overhead calculations
• proof