Machine learning based dynamic trust estimation framework for Securing wireless sensor networks.

Wireless Sensor Networks (WSNs) face significant challenges, such as security threats, and susceptibility to various internal and external attacks, impacting data integrity and network reliability. Trust-aware machine learning (ML) based security frameworks are vital solutions to overcome these challenges by improving security, decision-making, accuracy, and attack detection. These models utilizes ML to actively monitor trust levels, enhancing accuracy of malicious behavior detection and strengthening network security and integrity. In this context, we propose a Secure Machine-learning-based Adaptive Reliable Trust (SMART) model, suitable for unattended autonomous WSN environment. The proposed SMART model provides enhanced security and accuracy through fast and trustworthy dynamic generation of trust values. The proposed SMART model uses a novel ML algorithm that derives multiple trust features such as Co-Location Relationship (CLR), Co-Work Relationship (CWR), and Cooperativeness-Frequency-Duration (CFD). These features together give rise to an proactive trust score for sensor nodes, enabling predictions of likely misbehavior and allows for informed decisions regarding the reliability of individual sensor devices. Furthermore, SMART framework approach utilizes both immediate (direct) trust and derived (indirect) trust, making use of logical time windows to systematically watch for honest and suspicious interactions. SMART utilizes an efficient K-means clustering algorithm to assign data points to the nearest K-centers, optimizing cluster identification. Furthermore, It utilizes Principal Component Analysis (PCA) for significant variance identification and makes use of an SVM-based method for dimension reduction and accurate decision-making with lower time and space complexity. Performance measures such as accuracy, F1-score, recall, FNR, malicious SD detection, and trust value change measure its performance. Simulation results confirm that the SMART model can identify malicious nodes with a detection rate of 96%, FNR of 0.7%, F1-Score of 0.75, and accuracy of 96% even when there are 50 malicious nodes. This confirms the ability of the SMART model to enhance the trustworthiness and security of WSNs considerably.