SAHG: New Graph Model Achieves Top Accuracy Against LLM-Powered Social Bots

LLM-driven social bots now generate fluent, human-like text, making content-based detection alone ineffective. However, these bots leave relational patterns—interactions, behavioral similarity, and coordinated activity—that graph-based methods can exploit. Existing graph detectors struggle in two ways: Euclidean GNNs distort hierarchical social graphs, and fixed-curvature hyperbolic models assign uniform resolution to structural directions with different densities. Additionally, sophisticated bots forge heterophilic connections with real users, causing neighborhood aggregation to mix signals and dilute account-level evidence.

To address these challenges, the authors introduce SAHG. It learns a per-node, direction-dependent curvature field that adapts geometric resolution across structural directions. Sector prototypes convert angular concentration and alignment into classifier-readable features. Crucially, SAHG encodes per-account features and graph-neighborhood representations in two independent channels (the dual-channel design), fusing them only at the classifier to prevent contaminated aggregation. Evaluations on Fox8-23, BotSim-24, and MGTAB show SAHG achieves the highest accuracy and F1 across all benchmarks, outperforming feature-based, graph-based, LLM-based, and isotropic hyperbolic methods. Ablation and geometric analyses confirm the effectiveness of the anisotropic geometry and dual-channel setup.