Statistically Validated Benchmarking of Federated Learning Optimizers for Binary Intrusion Detection in Imbalanced IoMT Traffic

This work provides the first controlled, imbalance-aware comparison of five federated optimization algorithms (FedAvg, FedSGD, FedProx, FedAdam, and FedDyn) for binary intrusion detection in IoMT networks, using a single shared CNN-LSTM-ResNet backbone so that performance differences reflect the aggregation strategy alone. Evaluating each method under both IID and controlled label-skew non-IID client partitions on the CICIoMT2024 benchmark, we show that FedAvg delivers the strongest minority-class-aware performance in both settings, while FedDyn is most robust to label skew—revealing that the best optimizer depends on whether absolute performance or robustness is prioritized. We further demonstrate why accuracy is misleading under severe class imbalance (FedSGD reaches 98.996% accuracy yet fails entirely on the minority class), and we validate all findings with stratified paired bootstrap intervals and exact McNemar tests with Holm correction, alongside a matched centralized baseline on an identical frozen test set.