GATTrust: Learning Node Reputation in 6 G Networks Via a Hybrid GNN–SVM Framework

Trust evaluation is essential for secure cooperation in beyond-5 G/6 G networks, where decentralization, node mobility, and stringent latency constraints amplify security risks and complicate reliable decision-making. This study proposes GATTrust, a hybrid reputation inference framework that integrates Graph Attention Networks (GATs) for structure-aware representation learning with an SVM classifier using an Radial Basis Function (RBF) kernel for robust binary decision-making. We generated realistic node behavior traces in Network Simulator 3 (NS-3) and extracted trust-related features, including the interaction count, average delay, dropped packets, activity time, and synthetic trust scores. A sparse top-k graph is constructed using cosine similarity (with cosine-based edge weights), and a GAT encoder is trained to produce node embeddings that capture both the local interactions and global connectivity. The learned embeddings are reduced via Principal Component Analysis (PCA) and classified by the SVM with RBF kernel (RBF-SVM) to output the reputation labels and probability-calibrated trust scores. Experiments on large-scale graphs (up to 55k+ nodes and 1.6M edges) achieve a 97.97% validation accuracy and an F1-score of 0.99, outperforming representative trust baselines while remaining computationally practical. Overall, GATTrust provides a scalable and graph-aware mechanism for trust management and misbehavior detection in sixth-generation (6 G) infrastructure.