[Paper Review] A Novel Privacy-Preserved Recommender System Framework based on Federated Learning

Jiangcheng Qin, Baisong Liu|arXiv (Cornell University)|Nov 11, 2020

Privacy-Preserving Technologies in DataComputer Science56 references23 citations

TL;DR

This paper proposes a novel privacy-preserving recommender system framework (PPRSF) based on federated learning, enabling collaborative model training without centralizing user data. By decoupling global recall and local ranking models, it preserves user privacy, reduces communication costs, and maintains recommendation accuracy while complying with data protection regulations.

ABSTRACT

Recommender System (RS) is currently an effective way to solve information overload. To meet users' next click behavior, RS needs to collect users' personal information and behavior to achieve a comprehensive and profound user preference perception. However, these centrally collected data are privacy-sensitive, and any leakage may cause severe problems to both users and service providers. This paper proposed a novel privacy-preserved recommender system framework (PPRSF), through the application of federated learning paradigm, to enable the recommendation algorithm to be trained and carry out inference without centrally collecting users' private data. The PPRSF not only able to reduces the privacy leakage risk, satisfies legal and regulatory requirements but also allows various recommendation algorithms to be applied.

Motivation & Objective

• To address the critical privacy-personalization trade-off in recommender systems by eliminating centralized collection of user data.
• To design a scalable, privacy-preserving framework compliant with GDPR, CCPA, and other data protection regulations.
• To enable effective recommendation performance without compromising user data confidentiality or system efficiency.
• To support diverse recommendation algorithms within a secure, decentralized training architecture.

Proposed method

• Proposes a 4-layer hierarchical framework: Recall (server), Ranking (client), Re-Rank (server), and Service (client) layers.
• Decouples model training: global recall model is trained on server using public data, while local ranking models are trained on clients using private data.
• Employs federated learning with FedAvg-like aggregation: clients train locally on their private data and upload model updates to the server.
• Uses a two-stage process: server sends pre-filtered candidate items (Recall_Iik) to clients, who then rank them locally using their private data.
• Applies global model aggregation on the server to update the global ranking model, which is then redistributed to clients for local optimization.
• Separates public data (e.g., item features) from private data (e.g., user behavior), ensuring only non-sensitive data is shared.

Experimental results

Research questions

• RQ1 How can a recommender system maintain high personalization accuracy while ensuring user data is never centrally collected?
• RQ2 Can federated learning be effectively adapted to a multi-layered recommendation pipeline to preserve privacy?
• RQ3 How does the proposed framework balance privacy, communication efficiency, and recommendation performance?
• RQ4 To what extent can the framework comply with strict data protection regulations like GDPR and CCPA?
• RQ5 What architectural and training strategies are needed to support decentralized, privacy-preserving recommendation at scale?

Key findings

• The PPRSF enables training and inference without centralizing private user data, significantly reducing privacy leakage risks.
• The framework supports diverse recommendation algorithms through modular design, including content-based, collaborative filtering, and neural network models.
• Communication cost is reduced by pre-filtering item candidates on the server before sending them to clients.
• The system maintains recommendation accuracy through iterative global model aggregation using federated averaging.
• The architecture supports compliance with major privacy regulations by ensuring data remains on-device and is not shared with the server or other clients.
• The framework effectively mitigates both internal and external privacy threats by eliminating centralized data storage and access.