Threat adaptive trust weighted federated learning for Byzantine resilient cybersecurity across healthcare edge networks

The rising adoption of the Internet of Medical Things (IoMT) devices, edge computing, and distributed healthcare platforms brings with it a host of cybersecurity threats, whereas the centralized learning approach also brings with it privacy, communication, and single point of failure concerns. This study introduces the threat-adaptive Trust-weighted Federated Optimization (TrustFedHealth) framework for Byzantine-resilient healthcare cybersecurity, which combines the multi-signal client reliability estimation, temporal trust updating, norm clipping, error-feedback Top-k compression, and threat-guided Multi-Krum aggregation, with trust-weighted FedYogi optimization. The experiments were run on 15 k healthcare records with 52 attributes, with 12 k of them used for training and 3 k of them being future-held-out for testing, in 60 edge-node clients. Three tasks were assessed: security-risk classification, health-risk classification and health-risk regression. FedAdam showed the best security classification, with 93.67% accuracy, 0.9028 Macro-F1, 0.8867 MCC and 0.9898 ROC–AUC, while FedYogi attained the best ROC–AUC of 0.9904 and regression performance with 0.0516 MAE, 0.0651 RMSE and an R 2 of 0.6562. Multi-Krum has 0.8892 ROC–AUC and 75.77% security-classification accuracy when the attack rate is less than 20% malicious-client participation. A gender Macro-F1 gap of 0.0089 was achieved in the subgroup analysis and the expected calibration error was reduced from 0.0274 to 0.0185 by temperature scaling. The highest compressed checkpoint's accuracy is 90.46% and the highest Macro-F1 score is 0.8480, with a 22.25% cumulative communication reduction. The results show the promise of TrustFedHealth for secure, calibrated, equitable and communication efficient collaborative intelligence in the healthcare domain.