Lightweight AI Intrusion Detection System Targets IoT and Medical Device Threats

Researchers have unveiled a new intrusion detection system designed to protect Internet of Things (IoT) and Internet of Medical Things (IoMT) devices from increasingly sophisticated cyberattacks, reporting accuracy figures approaching 99.9 percent on two widely used benchmark datasets. The system, called AWFOL-IDS, short for Adaptive Weighted Feature-Optimized Lightweight Intrusion Detection System, was described in an open access paper published in Discover Internet of Things on 17 October 2026 by M. S. Padmini, V. Amruth, A. S. Sindhu, M. Asha Rani, V. K. Annapurna and C. S. Asha.

The motivation is stark. The authors note that distributed denial of service (DDoS) attacks have become among the most disruptive threats facing connected systems, with recent reports cited in the paper indicating that more than 20 million DDoS events were mitigated in a single quarter of 2025 and that attack bandwidth in large-scale incidents exceeded 30 terabits per second. Billions of malicious attempts per day are launched through automated botnets and compromised IoT devices, targeting critical infrastructure including healthcare, financial and industrial systems. As IoT botnets, AI-assisted hacking tools and multi-vector attack strategies proliferate, the researchers argue that traditional security mechanisms are no longer adequate.

Existing machine learning and deep learning based intrusion detection systems, despite reporting high detection accuracy, suffer from three persistent gaps identified by the team. Most methods treat feature selection, class-imbalance handling and classifier construction as independent, separately tuned stages, relying on static or dataset-agnostic feature subsets that cannot adapt when the balance between relevance and redundancy shifts across tasks or datasets. Where class imbalance is addressed, it is typically handled through generic sampling or loss-reweighting strategies such as the Synthetic Minority Over-Sampling Technique (SMOTE) applied uniformly across all features, meaning a minority-class signal carried by a weak, redundant feature is treated no differently than one carried by a highly discriminative feature. And techniques that achieve strong accuracy through deep or hybrid architectures incur high computational overhead, limiting their practicality for resource-constrained IoT and IoMT devices.

AWFOL-IDS addresses these gaps through an adaptive coupling of components. It integrates an Adaptive Weighted Information-Entropy Feature Selection (AWIE-FS) module with a Weighted Adaptive Random Forest with Feature-Integrated Optimization (WARF-FIO) classifier. The feature selection stage computes the mutual information between each feature and the class label to measure relevance, and pairwise mutual information between features to quantify redundancy. A dynamic weight for each feature balances its contribution to prediction against its redundancy with other features, with the weighting coefficients derived automatically from the dataset’s class-variance and feature-variance statistics rather than set by hand. A dynamic threshold, based on the mean and standard deviation of the feature weights, determines which features enter the final subset, so the selection adapts to the underlying distribution of feature importance.

The classifier extends conventional random forest in three ways. First, the adaptive feature weights from the selection stage are propagated directly into a weighted Gini index used for tree splitting, so features with higher discriminative power exert more influence during split selection. Second, class imbalance is handled through class-dependent sample weighting, in which each training sample receives a weight inversely proportional to its class frequency, and through weighted bootstrap sampling in which the selection probability of each sample is proportional to its weight. This increases the likelihood that minority class samples appear during training, improving their representation in each decision tree. Third, instead of letting all trees vote equally, the ensemble assigns performance-based weights to individual trees, with lower-error trees contributing more to the final prediction, and a regularization term penalizes excessively large feature weights to prevent over-reliance on a small subset of features.

The evaluation relied on two realistic benchmark datasets. CICIoT2023 represents a large-scale IoT network of 105 devices with 46,686,579 samples, of which 45,588,384 are attack samples and only 1,098,195 are benign, spanning 33 attack classes including Mirai, spoofing, brute force, web-based, reconnaissance, DoS and DDoS. CICIoMT2024 focuses on healthcare environments, built from a testbed of 40 devices, 15 simulated and 25 real, with 8,775,013 samples and 19 attack classes spanning spoofing, MQTT, reconnaissance, DoS and DDoS. Both datasets are severely imbalanced, with attack samples vastly outnumbering benign traffic, which is precisely the condition the new system is designed to handle.

The reported results are striking. On CICIoT2023, AWFOL-IDS achieved 0.9989 accuracy for binary classification, 0.9968 for 8-class and 0.9939 for 34-class tasks, with corresponding precision, recall and F-score values such as 0.9975 precision and 0.9968 recall for the binary case. On CICIoMT2024, it achieved 0.9998 accuracy for binary classification, 0.9992 for 6-class and 0.9856 for 19-class tasks. Stratified 5-fold cross-validation on the training partition showed better performance in terms of accuracy, precision, recall and F-score, and the hyperparameter selection was performed using cross-validation so that test-set results were not used at any stage of feature selection or tuning.

Generalization was tested in two ways. In cross-dataset experiments restricted to attack categories present in both datasets, namely DDoS, DoS, Reconnaissance and Spoofing, the system achieved 0.9924 accuracy when trained on CICIoT2023 and tested on CICIoMT2024, and 0.9891 accuracy in the reverse direction. The authors are careful to scope this claim: because the two datasets differ substantially in domain-specific attack types, the cross-dataset generalization applies only to the shared categories rather than the full attack surface of either environment. In an attack-family holdout experiment on CICIoT2023, in which the DDoS-PSHFlood family was excluded entirely from training, the system achieved 94.2 percent recall on the withheld family compared with 99.6 percent when it was included in training. A two-tailed paired t-test comparing per-fold F-scores against the strongest ablation baseline, a standard random forest with class balancing but without adaptive feature selection, yielded p equal to 0.0021, indicating the improvement is statistically significant at the 0.05 level.

An ablation study confirmed that each component matters. Removing feature selection dropped binary accuracy on CICIoT2023 to 0.9265, with further decreases in multi-class settings, while on CICIoMT2024 removal of feature selection reduced binary accuracy to 0.9352 and 19-class accuracy to 0.8724. Removing class balancing degraded recall and F-score especially in the 19-class case where imbalance is most severe, and the standard random forest baseline performed worst in all cases. Resource profiling showed that the feature selection stage reduces the feature space by 48.8 to 62.8 percent relative to the original feature set, lowering per-sample inference cost, with training time, latency, peak memory and model size measured on an Intel i3 processor with 8 GB of RAM and a 500 GB SSD.

The authors are candid about the limits of the work. The evaluation ran on a general-purpose CPU rather than representative resource-constrained IoT or IoMT edge hardware such as microcontrollers or embedded medical device processors, so deployment feasibility on genuinely constrained hardware remains to be validated. The study relies on two static benchmark datasets that represent a snapshot of attack behavior, with no evaluation against zero-day attacks, adversarially crafted traffic or concept drift, and no genuinely temporal or device-independent split was possible because the datasets lack reliable per-flow timestamps or device identifiers at the evaluated scale. The framework is supervised and depends on accurately labeled data, and it was evaluated offline rather than in a real-time streaming pipeline. Future work will focus on validating the system on constrained edge hardware, extending it with incremental learning, incorporating open-set recognition for zero-day detection, testing robustness against adversarial evasion, and exploring explainability techniques such as SHAP to support interpretability in clinical IoMT settings.

Subject of Research: A lightweight adaptive machine learning intrusion detection system for IoT and IoMT network security

Article Title: Adaptive lightweight intrusion detection system for internet of things and internet of medical things security

Article References: Padmini, M. S., Amruth, V., Sindhu, A. S., Rani, M. A., Annapurna, V. K., & Asha, C. S. (2026). Adaptive lightweight intrusion detection system for internet of things and internet of medical things security. Discover Internet of Things, 6(1), Article 215. https://doi.org/10.1007/s43926-026-00531-3

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00531-3

Keywords: intrusion detection, Internet of Things, Internet of Medical Things, machine learning, random forest, class imbalance, feature selection, DDoS, cybersecurity, CICIoT2023, CICIoMT2024, healthcare security