PrivFR: Privacy-Enhanced Federated Recommendation With Shared Hash Embedding

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Federated recommender systems (FRSs) offer privacy benefits by training models on distributed user data.
• Conventional FRSs face challenges with high carbon footprints due to large model transmissions and the out-of-vocabulary (OOV) item issue.

Purpose of the Study:

• To propose an efficient and privacy-enhanced federated recommendation framework for cross-device settings.
• To address the limitations of current FRSs regarding resource consumption and handling OOV items.

Main Methods:

• Introduced PrivFR, a framework utilizing shared hash embedding and multiple hash functions for resource-efficient item representation.
• Developed a mechanism to maintain a small shared hash embedding pool on local clients, avoiding storage of all item embedding vectors.
• Provided theoretical proof of data privacy protection for local clients.

Main Results:

• PrivFR significantly reduces storage and communication overheads compared to traditional FRSs.
• The proposed method effectively handles the OOV item problem in federated recommendation scenarios.
• Experimental results demonstrate that PrivFR outperforms existing state-of-the-art FRSs.

Conclusions:

• PrivFR offers a novel solution for efficient and privacy-preserving federated recommendation.
• The shared hash embedding approach is effective in mitigating resource constraints and addressing the OOV issue in FRSs.
• This framework enhances the practicality and performance of FRSs in real-world cross-device applications.