Stronger Privacy for Federated Collaborative Filtering with Implicit Feedback

Recommender systems are commonly trained on centrally collected user\ninteraction data like views or clicks. This practice however raises serious\nprivacy concerns regarding the recommender's collection and handling of\npotentially sensitive data. Several privacy-aware recommender systems have been\nproposed in recent literature, but comparatively little attention has been\ngiven to systems at the intersection of implicit feedback and privacy. To\naddress this shortcoming, we propose a practical federated recommender system\nfor implicit data under user-level local differential privacy (LDP). The\nprivacy-utility trade-off is controlled by parameters $\\epsilon$ and $k$,\nregulating the per-update privacy budget and the number of $\\epsilon$-LDP\ngradient updates sent by each user respectively. To further protect the user's\nprivacy, we introduce a proxy network to reduce the fingerprinting surface by\nanonymizing and shuffling the reports before forwarding them to the\nrecommender. We empirically demonstrate the effectiveness of our framework on\nthe MovieLens dataset, achieving up to Hit Ratio with K=10 (HR@10) 0.68 on 50k\nusers with 5k items. Even on the full dataset, we show that it is possible to\nachieve reasonable utility with HR@10>0.5 without compromising user privacy.\n