TH-GNN Detects LLM-Agent Shilling Attacks via Temporal Graphs

TH-GNN uses a heterogeneous temporal graph neural network with a Heterogeneous Graph Transformer backbone to detect LLM-generated fraudulent reviews. The model fuses RoBERTa-based semantic embeddings with structural user embeddings and GRU-based temporal analysis to capture coordination patterns that text-only detectors miss.