Abstract
Decision trees are increasingly favored in lightweight Network Intrusion Detection Systems (NIDS) for Medical IoT (MIoT) due to their efficiency and interpretability, which are critical for resource-constrained healthcare environments. However, these models remain vulnerable to adversarial attacks that can manipulate network traffic features and distort decision boundaries, resulting in dangerous misclassifications of malicious activity. Existing defense mechanisms often lack robustness, particularly when confronted with unseen attack strategies. To address these challenges, we introduce the Manifold-based Defense framework (M4D), tailored for decision tree-based NIDS in MIoT. M4D entangles projected latent spaces to enforce interdependencies among traffic features, thereby impeding low-cost adversarial manipulation. Additionally, it incorporates cross-manifold mixup augmentation to further smooth decision boundaries and enhance robustness against adversarial perturbations. Extensive evaluation on two recent MIoT network intrusion datasets, CICIoMT2024 and ICU, under both benign and five diverse adversarial attack scenarios, demonstrates that M4D consistently maintains over 90% accuracy, surpassing six benchmark defense methods. These findings underscore the potential of M4D for enhancing the security and reliability of anomaly detection in critical medical IoT networks.