Explainable Artificial Intelligence-Driven Zero-Trust Architecture for Secure and Trust-Aware Decentralized IoT Networks

Abstract

The widespread integration of Internet of Things (IoT) technology for environmental monitoring systems has produced a decentralized data environment where various types of devices continuously acquire and share sensing data. Despite providing significant data availability and scalability, it also exposes various security challenges, including unauthorized access, malicious information sharing, and ineffective selection of trustworthy peers. Traditional security models often rely on predefined authentication mechanisms and attack detection, but this model may not adaptively modify the trust assigned to a device when its behavior is mistrustful. These challenges necessitate the need for a dynamic trust evaluation mechanism to regularly monitor and evaluate the trustworthiness of IoT peers’ devices. Therefore, this study introduced an AI-Driven Zero-Trust Peer-to-Peer Security Framework (AIZT-P2PSF), aiming to develop a reliable and secure model that effectively classifies attack detection and dynamic trust evaluation. To achieve this, the model utilizes the TON-IoT dataset, including seven heterogeneous IoT devices; a unique device_ID is assigned to each individual device. The pre-processed data are forwarded to feature engineering and temporal sequences to extract behavioral, environmental, and temporal features and construct device-specific temporal sequences for learning peer behavior over time. Then, the CNN–BiLSTM–attention model is employed to accurately recognize the attack pattern and produce an attack probability. Furthermore, a dynamic trust estimation module is integrated to determine the trust score using historical pattern, cyber risk, and environmental consistency. Finally, the updated trust score enables dynamic zero-trust decisions and selects trusted peers. The experimental evaluation was carried out through the TON-IoT dataset across various evaluation metrics; the findings highlight the effectiveness of the proposed framework over existing techniques by demonstrating 98.74% accuracy and validate its potential in accurately categorizing attack detection, trust management, and peer selection of decentralized peer-to-peer IoT networks.