A federated learning-based dual-filtering solution to combat data poisoning attacks in medical image classification

Abstract The deep integration of artificial intelligence with the Internet of Medical Things (IoMT) has accelerated the adoption of federated learning (FL) for medical image classification. However, the decentralized training pipeline exposes FL to poisoning threats at both the data and model-update levels, potentially compromising diagnostic reliability. To address this challenge, we propose FedGF, an end-to-end two-stage defense scheme that combines client-side Genetic Data Selection (GDS) with server-side Federated Unlearning (FU) enhanced by LOF-based anomaly detection. GDS optimizes local training subsets through a fitness function that jointly considers parameter similarity, validation accuracy, and category-bias penalty, thereby filtering potentially poisoned samples before local training. During aggregation, FU further identifies and removes anomalous client updates and performs similarity-aware robust weighting. Experiments on the COVID-19 Radiography Database and the Pneumonia Chest X-ray Dataset under label-flipping attacks demonstrate that FedGF consistently improves robustness and convergence stability compared with conventional aggregation defenses.