Building a privacy-preserving Federated Recommender system for mobile devices

A new federated recommendation pipeline for mobile devices separates non-sensitive preference data from sensitive on-device context, keeping personal signals local while still enabling personalized content delivery, according to a paper published on arXiv [1]. The two-stage system runs a collaborative filtering model on non-sensitive app-context data in the cloud to generate a shortlist of relevant items, then re-ranks those candidates directly on the device using sensitive mobile signals [1]. Only model updates and gradients leave the device, a design the authors describe as a response to the tension between personalization and privacy regulations that restrict centralized data pooling [2]. The approach was validated on the MovieLens dataset, the UCI Human Activity Recognition dataset, and a proprietary pilot dataset [1]. A production-ready implementation is provided as a Kotlin Multiplatform library deployable on both Android and iOS [2]. The work arrives as mobile platforms that rely on recommendation algorithms face sustained scrutiny over data handling. YouTube, which Google acquired in 2006 for $1.65 billion, has drawn criticism for recommendation algorithms that have been shown to promote conspiracy theories and falsehoods [4]. Google Maps, used by more than one billion people monthly as of 2020, collects real-time location and traffic data to power its services, illustrating the scale of sensitive mobile-context data that centralized systems routinely process [3]. The concept of digital self-determination has emerged as a framework to address the erosion of individual agency that accompanies the digitization of daily life, emphasizing the need for architectures that preserve user autonomy [5]. By keeping sensitive context data on-device and only transmitting model gradients, the proposed pipeline aligns with that principle while maintaining the utility of personalized recommendations [1][2].