FedFly: Towards Migration in Edge-based Distributed Federated Learning
Due to mobility, a device participating in Federated Learning (FL) may disconnect from one edge server and will need to connect to another edge server during FL training. This becomes more challenging when a Deep Neural Network (DNN) is partitioned between device and edge server referred to as edge-based FL. Moving a device without migrating the accompanying training data from a source edge server to the destination edge server will result in training for the device having to start all over again on the destination server. This will in turn affect the performance of edge-based FL and result in large training times. FedFly addresses the mobility challenge of devices in edge-based distributed FL. This research designs, develops and implements the technique for migrating DNN in the context of edge-based distributed FL.
FedFly is implemented and evaluated in a hierarchical cloud-edge-device architecture on a lab-based testbed to validate the migration technique of edge-based FL. The testbed that includes four IoT devices, two edge servers, and one central server (cloud-like) running the VGG-5 DNN model. The empirical findings uphold and validates our claims in terms of training time and accuracy using balanced and imbalanced datasets when compared to state-of-the-art approaches, such as SplitFed. FedFly has a negligible overhead of up to 2 seconds but saves a significant amount of training time while maintaining accuracy.
- AM (Accepted Manuscript)
CitationUllah , R , Wu , D , Harvey , P , Kilpatrick , P , Spence , I & Varghese , B 2022 , ' FedFly : towards migration in edge-based distributed federated learning ' , IEEE Communications Magazine , vol. 60 , no. 10 , pp. 42-48 . https://doi.org/10.1109/mcom.003.2100964
Cardiff Met Affiliation
- Cardiff School of Technologies
Cardiff Met AuthorsRehmat Ullah
- © The Publisher
Publisher Rights StatementCopyright © 2022 IEEE. This work has been made available online in accordance with publisher policies or with permission. Permission for further reuse of this content should be sought from the publisher or the rights holder. This is the author created accepted manuscript following peer review and may differ slightly from the final published version. The final published version of this work is available at https://doi.org/10.1109/mcom.003.2100964