A Reliable Federated Learning Approach in Edge Computing

Abstract

The federated learning scheme proposes keeping data at the edge nodes, and bringing a key part of model training to the edge. So, in federated learning, edge parties maintain their own data and train the model in a distributed manner. Gradients or model updates are taking place between the edge participants and the centric-aggregator. Therefore, there is need to improve the processes that compose the federated learning training steps so that the system will gain better selection for participants in the process and apply an appropriate aggregation strategy that lets the global trained model's accuracy improve and takes advantage of the data diversity, especially in heterogeneous environments that are more common in reality. Adopting a federated learning system requires consideration of potential negative aspects. As long as a federated learning system more probably operates in an open environment that belongs to different organizations or individuals, that means the system is subjected to both honest participants that work properly and do the right duties in the process well, and dishonest participants that intentionally desire to affect the system and degrade the FL performance or get advantage from the system information. A reputation-based scheme for selecting trustworthy end nodes to participate in the federated learning process to be selected mainly in terms of measuring dataset overlaps, beside the embedded reputation that resulted from the contributions and integrity of participants in previous tasks with the owner, can be combined as an additional reputation term in case other cooperation between task owner and participant has taken place before

Keywords: Machine learning, Edge computing, Global model, Horizontal Federated learning