IRAB-NET: An Interpretable Hybrid Boosting Framework for Intelligent Routing Attack Detection and Reliability Estimation in WSNs

Abstract

Wireless Sensor Networks (WSN) are widely applied in environmental monitoring, healthcare, and industrial automation, where secure and reliable communication is critical. However, WSNs are highly susceptible to routing attacks such as Sybil, Sinkhole, and Selective Forward, which degrade network performance and compromise data integrity. The main challenge is accurate attack detection while simultaneously evaluating node reliability through trust and reputation scores. Traditional approaches based on standalone Machine Learning (ML) models or rule-based techniques lack adaptability, fail to capture complex nonlinear relationships, and perform poorly in dynamic environments, resulting in limited accuracy and generalization. To overcome these limitations, this research proposes an advanced framework that leverages ML techniques on blockchain-based dataset characteristics for intelligent routing attack analysis. The system is implemented using the Django framework, enabling secure, role-based access for Network AI Engineers and WSN Engineers, supporting both prediction and batch analysis functionalities. Multiple models are utilized, including K-Nearest Neighbors (KNN), Passive Aggressive (PA), AdaBoost (AB), and a hybrid one classification and two regression trees (1CA2RT). A novel Interpretable Recurrent Gradient-based Adaptive Boosting Network (IRAB-NET) is introduced, where Gradient Boosting (GB) enhances classification of node behavior, and AB improves regression of trust and reputation scores. Recurrent Neural Network (RNN) components act as base estimators to capture temporal and complex patterns. A structured preprocessing pipeline ensures effective feature engineering, encoding, and normalization. The proposed model achieves 99.44% classification accuracy with reduced regression errors, enabling robust attack detection and reliable decision-making in WSN environments.