Detecting Free-Riders in Federated Learning Using an Ensemble of Similarity Distance Metrics

Federated learning (FL) has witnessed an increase in adoption mainly because of its practical implications and ability to leverage large amounts of data while safeguarding their privacy. However, at its core, FL highly relies on the resources of its participants (i.e., clients), that directly influence the accuracy of the global model in any given training task. A significant challenge arises from the presence of clients that do not fully engage in the training process, yet still benefit from the updated models provided by the server. This lack of active participation has the potential to undermine the overall performance of the global model. Furthermore, free-riders create an unfair scenario where honest participants contribute more while receiving the same overall benefit. In this work, we propose a strategy that relies on an ensemble of several distance measures to mitigate the impact of free-riders in FL environments. By integrating multiple distance metrics into a unified ensemble approach, our objective is to detect and identify free-riders effectively. Extensive simulations and experimental results highlight the robustness of our approach.