Dynamic Weighted K-Asynchronous Federated Learning for Privacy-Preserving Intrusion Detection in Wireless Body Area Networks

Wireless Body Area Networks (WBANs) are critical enablers of real-time, patient-centric healthcare, yet their open wireless interfaces make them prime targets for cyberattacks, including spoofing, DoS, and data tampering, that can compromise both patient safety and medical data integrity. Centralized intrusion detection systems (IDS) are fundamentally unsuitable for WBANs due to privacy violations from raw bio-data transmission, high communication overhead, and energy inefficiency. To overcome these challenges, we propose DW-KAFL: a novel Dynamic Weighted K-Asynchronous Federated Learning framework that redefines secure edge intelligence in healthcare IoT. DW-KAFL integrates four synergistic innovations: (1) a K-asynchronous update protocol that eliminates straggler bottlenecks by aggregating from any K responsive nodes per round; (2) dynamic trust-aware weighting using a composite score $$w_i = \alpha \cdot \textrm{acc}_i + \beta \cdot \mathrm {bio\_sim}_i + (1-\alpha -\beta )\cdot E_i^{\text {norm}}$$ , which prioritizes accurate, physiologically consistent, and energy-sustainable nodes; (3) a hybrid CNN-LSTM local IDS enhanced with a VAE-based Bio-Anomaly Fusion Layer, enabling detection of stealthy attacks that decouple network anomalies from expected physiological responses; and (4) end-to-end $$(\epsilon ,\delta )$$ -differential privacy ( $$\epsilon _T=1.0$$ ) to ensure HIPAA-compliant model updates. We evaluate DW-KAFL on a realistic fused dataset combining CIC-IDS2018 and MIMIC-III, under non-IID conditions and adversarial model poisoning. Results show that DW-KAFL achieves an intrusion detection accuracy of 97.5% and an F1-score of 0.968, outperforming state-of-the-art baselines including FedAvg, PAG-FL, and FL-SCNN-Bi-LSTM. It reduces convergence latency by 28%, cuts communication energy by 35%, and maintains robustness with only a 4.9% accuracy drop under 30% malicious nodes. Theoretical analysis establishes O(1/T) convergence under bounded staleness and sparse composition. By fusing machine learning with physiological context, DW-KAFL sets a new benchmark for secure, efficient, and clinically intelligent e-health systems.