Making resource adaptive to federated learning with COTS mobile devices

Mobile devices are pervasive data producers that bridges users and emerging network services, e.g., learning techniques. Today, mobile devices are continuously generating user related data, which at most time is privacy concerned, at the edge of network. Indeed, such privacy concerned and plentiful data is naturally in-depth coupled with modern distributed learning paradigms, e.g., federated learning. However, modern distributed learning paradigms even those based on cloud computing can still result in heavy computation burden on resource constrained mobile environments. In this paper, we tackle the challenge arising from modern distributed learning with resource constrained mobile devices by no data delivery. To this end, this paper proposes MobiFed, a resource adaptive distributed learning system for mobile scenarios to address the resource heterogeneity issue in commercial off-the-shelf (COTS) mobile federated community. Experimental results demonstrate that MobiFed not only reduces the system time overheads to achieve the promised learning accuracy (by up to \(47\%\)), but also improves the quality of global federated learning system, e.g., almost \(10\%\) and even higher accuracy than the promised performance in comparing with existing other systems. In addition, MobiFed provides a user friendly and self-managed resource mechanism, is tolerant for computation fault recovery, and provides excellent extensibility for potential mobile device’s plugging-in operations.