Robust dynamic trust evaluation using snapshot-based graph neural networks in on-line social networks

Résumé fourni par la source

Trust evaluation is a fundamental component of decision making in online social networks, as it analyzes user interactions and identifies malicious behaviors. While recent graph neural network (GNN)–based models improve trust modeling, they remain vulnerable to adversarial attacks because they rely on fixed temporal segmentation and lack explicit awareness of when structural disruptions occur. This paper proposes TEAMS, a robust trust evaluation model that integrates adaptive snapshot segmentation with spatio-temporal GNN architectures. The proposed segmentation layer jointly analyzes multiple structural indicators including graph density, degree statistics, edge growth, and gradient-based change signals to detect meaningful change points and construct snapshots centered on critical structural events. This event-driven segmentation reduces redundant temporal information and enables more precise modeling of attack-prone periods. On top of these snapshots, TEAMS employs a dual-role spatial aggregation layer that explicitly models trustor–trustee asymmetry using reliability-aware neighbor selection, along with a position-aware temporal attention layer that captures the evolution of trust relationships over time. By focusing learning on structurally significant transitions rather than uniform time intervals, TEAMS enhances robustness against collaborative good-mouthing, bad-mouthing, and adversarial edge injection attacks, including scenarios involving unseen nodes. Extensive experiments on the real-world Bitcoin-OTC and Bitcoin-Alpha trust networks demonstrate that TEAMS achieves the strongest MCC and F1-macro performance in most evaluation settings, with up to 26.88% improvement in MCC and 8.68% improvement in F1-macro over the strongest competing baseline in the single-timeslot setting, while remaining competitive in AUC and BA. Moreover, TEAMS substantially recovers trust prediction performance under various attack scenarios, confirming its robustness to malicious behaviors in dynamic trust networks.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.