Researchers propose a new method called TH-GNN to detect fake reviews generated by large language models (LLMs). These fake reviews can be hard to spot because they are realistic and coordinated. The proposed method combines graph structure, temporal patterns, and semantic information to identify suspicious activity. It uses a transformer-based backbone and attention mechanisms to analyze user behavior and review content. In experiments, TH-GNN achieved high detection accuracy across various attack scenarios and datasets. --- Why it matters: This matters because sophisticated LLM-driven shilling attacks can compromise recommender systems, leading to biased or fake recommendations. Engineers working on AI-powered recommendation systems need effective methods to detect such attacks and maintain the trustworthiness of their systems. Source: https://arxiv.org/abs/2608.20376