New TH-GNN model detects LLM-agent shilling attacks

Researchers have developed TH-GNN, a novel heterogeneous temporal graph neural network designed to detect sophisticated shilling attacks orchestrated by LLM agents. This model utilizes a two-layer Heterogeneous Graph Transformer backbone with attention mechanisms and temporal encodings to analyze both the semantic content of reviews and the structural and temporal patterns of user interactions. By jointly modeling these signals, TH-GNN significantly outperforms existing text-only detection methods, achieving a grand-mean F1 score of 0.870 across various attack scenarios. AI

IMPACT This research offers a new method for detecting sophisticated LLM-driven attacks on recommender systems, potentially improving platform integrity.

RANK_REASON Academic paper detailing a new model and its evaluation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Agent4SR
• arXiv
• gated recurrent unit
• Hugging Face
• LLM agents
• Roberta
• TH-GNN

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →