Towards Blockchain-Based Reputation-Aware Federated Learning
Muhammad Habib ur Rehman, Khaled Salah, Ernesto Damiani, Davor Svetinović
Abstract
Federated learning (FL) is the collaborative machine learning (ML) technique whereby the devices collectively train and update a shared ML model while preserving their personal datasets. FL systems solve the problems of communication-efficiency, bandwidth-optimization, and privacy-preservation. Despite the potential benefits of FL, one centralized shared ML model across all the devices produce coarse-grained predictions which, in essence, are not required in many application areas involving personalized prediction services. In this paper, we present a novel concept of fine-grained FL to decentralize the shared ML models on the edge servers. We then present a formal extended definition of fine-grained FL process in mobile edge computing systems. In addition, we define the core requirements of fine-grained FL systems including personalization, decentralization, fine-grained FL, incentive mechanisms, trust, activity monitoring, heterogeneity and context-awareness, model synchronization, and communication and bandwidth-efficiency. Moreover, we present the concept of blockchain-based reputation-aware fine-grained FL in order to ensure trustworthy collaborative training in mobile edge computing systems. Finally, we perform the qualitative comparison of proposed approach with state-of-the-art related work and found some promising initial results.