ADAPTIVE ERROR-AWARE LIGHT GRADIENT BOOST MACHINE FOR CROSS-DATASET INTRUSION DETECTION IN MANETS

Abstract

The purpose of this article is to develop an adaptive error-aware light gradient boost machine for cross-dataset intrusion detection in Mobile Ad hoc Networks (MANETs). MANETs are dynamic in nature and thus vulnerable to various attacks, such as black hole, grey hole, and tapping attacks. Therefore, to mitigate security issues, intrusion detection in MANETs increasingly relies on machine learning models, with existing studies dominated by accuracy-centric evaluation. However, there are still limitations of such metrics in highly separable and imbalanced intrusion datasets. Therefore, this study proposes an Adaptive Error-Aware LightGBM (AEA-LGBM) framework for intrusion detection across datasets. The methods used include preprocessing the data through encoding and feature standardisation, as well as training and testing. The framework dynamically adjusts instance weights to minimise the false negative rate (FNR) while integrating an error-efficiency evaluation strategy that combines FNR, false positive rate (FPR), and inference time. Experiments using CIC-DDoS2019 and CICIDS2017 datasets show that AEA-LGBM reduces FNR from 0.000273 to 0.000068 during internal evaluation while maintaining near-perfect accuracy. Independent cross-dataset validation reveals substantial performance degradation across conventional models due to domain shift. To address this limitation, a transfer-aware heterogeneous learning strategy is introduced, achieving 99.90 per cent accuracy, 99.94 per cent recall, and 0.000580 FNR under unseen cross-environment evaluation. Therefore, the model effectively minimises the false negative rate through adaptive weighting, thereby improving MANET security.