FedLC-Trans: privacy-preserving CNN–Transformer fusion for resilient IoMT intrusion detection

Abstract

The objective of this study is to develop an accurate and computationally efficient intrusion detection framework for distributed Internet of Medical Things (IoMT) environments while reducing the need to centralize raw client data. We propose FedLC-Trans, which integrates MobileNet-based lightweight convolutional blocks for local feature extraction, an Efficient Transformer for modeling interactions and dependencies among feature-token representations, and an attention-guided SeqPool module for discriminative feature aggregation. Federated Averaging (FedAvg) is employed to collaboratively train the model across five non-IID clients without transferring raw client data to the central server. To reduce the risk of evaluation leakage, normalization, Random Forest-based feature selection, PCA, and SMOTE are derived or applied using the training data only, while the learned transformations are subsequently applied to the held-out test data. FedLC-Trans was evaluated on the WUSTL-EHMS, ECU-IoHT, and CICIoMT2024 datasets and achieved accuracies of 0.9913, 0.9892, and 0.9916, respectively. The results indicate consistently high intrusion-detection performance across the three evaluated IoMT datasets. However, computational profiling was performed on an NVIDIA A100 GPU rather than on resource-constrained IoMT edge hardware, and federated learning alone does not provide formal protection against gradient leakage, model inversion, or membership-inference attacks. Therefore, the present findings demonstrate reduced centralized raw-data exposure and promising computational characteristics, while edge-device validation and stronger privacy-preserving mechanisms remain important directions for future work.