Peer to Peer Federated Learning: Towards Decentralized Machine Learning on Edge Devices
Aristeidis Karras, Christos Karras, Konstantinos C. Giotopoulos, Dimitrios Tsolis, Κωνσταντίνος Οικονόμου, Spyros Sioutas
Abstract
Federated Learning (FL) is an emerging technique that assures user privacy and data integrity in distributed machine learning environments. To perform so, chunks of data are trained across edge devices and a high performance cluster server maintains a local copy without exchanging it with other parties. In this work, we investigate a FL scenario in a real-world case study using 5, 10 and 20 Raspberry Pi devices acting as clients. Under this setup, we employ the widely known FedAvg algorithm which trains each client for several local epochs and then the weight of each model is aggregated. Moreover we perform experiments across imbalanced and noisy data so as to explore scalability and robustness based on real-world datasets were noise is present and we also propose two innovative algorithms where the FL scenario is considered as a peer-to-peer formulation. Ultimately, to ensure that each device is not oversampled a client-balancing Dirichlet sampling algorithm with probabilistic guarantees is proposed.