Collaborative trust computation model for social internet of things

Abstract

Trust compromise is a major problem in social internet of things (SIoT). Existing trust management approaches often rely on predefined trust values, limited behavioral attributes, or historical interactions, making them less effective under dynamic and cold-start environments. This paper proposes a deep collaborative event trust model (DCETM) integrating collaborative event confidence modeling with graph neural networks (GNNs), long short-term memory (LSTM) networks, and multi-attribute trust evaluation to estimate device trustworthiness from spatial, temporal, and behavioral information. The proposed framework derives evidence-driven trust scores from collaborative event observations and integrates reliability, consistency, response behavior, and event confidence for robust malicious node detection. A preference-aware topology refinement mechanism further isolates malicious devices and promotes trustworthy interactions. Experimental evaluation on a SIoT benchmark dataset demonstrates that the proposed DCETM achieves 93.4% attacker detection accuracy, 0.90 F1-score, and reduces the misclassification rate by 25–30% compared with recent deep learning-based trust models. In-addition proposed DCETM maintains reliable trust estimation under cold-start conditions by avoiding default trust assignment through collaborative event validation, demonstrating its suitability for secure and scalable SIoT environments.