Using attentive temporal GNN for dynamic trust assessment in the presence of malicious entities

Abstract

Trust networks are built on the premise that there is a network of trust among entities. Precisely evaluating pairwise trust in such networks is essential for enhancing the quality of service in various applications, including online marketing, recommender systems, and IoT-based applications. However, it remains challenging to consider dynamic trust and account for malicious behavior in a scalable and effective manner.Recently, Graph Neural Networks (GNN)-based methods have emerged as powerful and scalable tools for trust estimation. Nevertheless, they lack thorough consideration for the time-dependent nature of trust, especially in the presence of malicious behaviors, leading to inaccurate trust predictions. To address this gap, we propose a novel trust assessment model called MATA, which handles the time dependency of trust and exhibits robustness against malicious behaviors. MATA leverages an extended Graph Convolutional Network (GCN) integrated with Recurrent Neural Networks (RNNs) and an attention mechanism to jointly capture time dependencies and trace entities’ behavior over time in complex networks. Additionally, by clustering the learned trust representations over time, we assign reputation values to entities. This allows us to monitor their dynamic behavior and mitigate potential adverse effects on trust assessment.In our experiments on real-world datasets, MATA outperforms the state-of-the-art methods by up to 8.4 % in accuracy. This improvement is a testament to the effectiveness of the proposed approach. Furthermore, our work enhances the robustness of trust assessment models, as evidenced by consistently higher performance metrics in the face of various trust-related attacks.