Fedai: Federated recommendation system with anonymized interactions

Federated learning has proved its success in recommendation services as federated recommendation system, in which the features of users and items can be learned locally without uploading sensitive personal interaction records for privacy concerns. However, in the context of large-scale service, huge system cost is generally inevitable to accommodate massive local models trained by diverse user behaviors and preference. Furthermore, privacy-preserving methods have to bring in noisy perturbation or complex encryption schemes during local model aggregation process, which could make the system efficiency and scalability issues even worse. In this paper, we are motivated to present a system design for large-scale recommendation service, in which the high-order interactions are utilized to balance the federated model training, user privacy preservation and system scalability. Specifically, a homomorphic encryption based grouping strategy is designed to scale down the federated model training and extract high-order interactions for group members without privacy leakage. Then we present an attention-based mechanism to leverage the inner group high-order interactions for local user embedding enhancement. During the global item embedding aggregation, user private data can be protected by local interaction anonymization without extra noise perturbation involvement. Experimental results on five datasets demonstrate the effectiveness of our method while ensuring user data privacy.