A novel heterogeneous temporal graph neural network, TH-GNN, is designed to detect LLM-driven shilling attacks in recommender systems, achieving a grand-mean F1 score of 0.870 across five attack families and four datasets.

The study emphasizes that integrating temporal, structural, and semantic signals closes the detection gap left by single-modality detectors.

Cross-modal fusion enables comparing a user’s graph behavior with their reviews to detect mismatches, strengthening detection beyond single-modality approaches.

Temporal burstiness modeling captures coordinated bursts typical of shilling campaigns, improving detection in low-activity conditions.

Key architectural components include learnable sinusoidal temporal encodings on edges, cross-modal attention using frozen RoBERTa embeddings for semantic fusion, and GRU-based temporal burstiness modeling.

LLM agents can generate credible shilling profiles, fluent reviews, and coherent ratings at scale, challenging defenses that target simpler manipulation patterns.

TH-GNN fuses graph structure, temporal signals, and language content through a two-layer heterogeneous graph transformer with per-type and per-relation attention.

The article provides a FAQ addressing why LLM agents challenge defenses, the limitations of detectors, and the performance gains of TH-GNN.

Existing detectors fail because text-only methods miss graph structure and timing patterns, while graph-only methods miss review semantics and cross-modal inconsistencies.

Evaluation across diverse attack families and datasets shows robust performance, with TH-GNN outperforming text-only baselines by 10.9 percentage points on Agent4SR attacks and by 11.5 points at the lowest injection rate.