RepuFL: reputation-aware client selection in Federated Learning for mobile networks

The mounting demand for data processing in mobile networks, together with the inherent limitations of mobile devices, has as of late prompted the use of the Federated Learning (FL). FL is a distributed machine learning paradigm for distributing and aggregating a model's parameters while maintaining the data privacy. However, corrupt mobile devices, i.e., also referred to as participants or clients in the FL, are capable of interfering with the FL training process primarily owing to its lack of control over the local model training and which may eventually jeopardize the integrity of the global aggregation. This research therefore adopts a reputation-aware approach for quantifying the credibility of the participants for the FL tasks. The envisaged approach considers adaptive and dynamic parameters, i.e., historical accuracy, data quality, willingness, and behavioral consistency, in a bid to quantify a participant's reputation. In this work-in-progress paper, we present the foundational concept and framework of our envisaged model together with the preliminary experimental evaluations. It also include details on the datasets used, evaluation metrics, and a comparative analysis with a state-of-the-art method. A more comprehensive investigation of the FL training process together with an in-depth performance evaluation would be addressed in our future work.