Towards More Robust Federated Learning with Medical Imaging Model Anomaly Detection

Federated Learning (FL) is a decentralized approach to Machine Learning (ML) that allows distributed model training locally across different devices, instead data transportation to the aggregation point, thus increasing the data privacy protection. Nevertheless, adversarial attacks still possess a problem for FL in real-world applications, aiming to deteriorate the aggregated ML model performance or reconstruct the training data using the model updates. In this paper, we prove that our Model Anomaly Detection (MAD) technique based on model trust evaluation not only allows for the robust identification of the malicious participants during the distributed ML model training on early aggregation stages, but also improves the model convergence. We verify our findings on the subset of the MedicalMNIST dataset.