Detecting LLM-Generated Spam Reviews by Integrating Graph Neural Network and Language Model Embeddings

ACL ARR 2025 February Submission6128 Authors

Abstract: Detecting spam reviews has drawn much attention for years. Many efforts have been dedicated to detecting deceptive spam reviews, accumulating rich literature and plentiful effective practices. However, the recent rapid development of large language models (LLMs) brings new challenges to this area. Fraudsters could misuse LLMs to write highly authentic and misleading fake reviews. To detect such harmful contents, we formulate the detection as a node classification task on the constructed review graph and employ the graph neural network (GNN) to handle the users' behavior. More specifically, we seamlessly integrate gated graph transformers with the language model to embed the review texts where previously engineered features summarized by fraud experts are insufficient. The Experiments show that this integration in our method FraudSquad turns out to be effective on two created LLM-attacked and two human-attacked spam review datasets, outperforming state-of-the-art detection methods. Moreover, FraudSquad achieves a modest model size and requires very few training labels, making defending the spam review attack more practical.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: rumor/misinformation detection

Contribution Types: Model analysis & interpretability, Approaches to low-resource settings, Publicly available software and/or pre-trained models, Data resources

Languages Studied: English, Chinese

Submission Number: 6128