10th U6+ Consortium of African Universities International Conference

Mobile Ad Hoc Networks (MANETs) are highly vulnerable to cyberattacks due to their decentralized architecture, dynamic topology, and resource constraints. Existing machine learning-based intrusion detection systems often prioritize overall classification accuracy while overlooking false negative errors that allow attacks to remain undetected. This paper proposes an Adaptive Error-Aware Light Gradient Boosting Machine (AEA-LGBM) framework that integrates adaptive sample weighting, error amplification, adaptive threshold optimization, and SHAP-based feature optimization to minimize missed attacks while maintaining computational efficiency. The framework was evaluated on a realistic NS-3-generated MANETs dataset containing 39,036 flow records using a leak-free 70%–15%–15% train–validation–test protocol and compared with conventional machine learning models and LightGBM variants using both conventional and error-sensitive performance metrics. AEA-LGBM reduces the False Negative Rate from 9.42% to 0.56%, 99.44% recall and maintaining fast inference performance. Therefore, the framework is suitability for deployment-oriented intrusion detection in resource-constrained MANETs environments.