A Federated Multi-Modal Graph Neural Network Framework For Scalable and Privacy-Aware Fake Review Detection In E-Commerce

Abstract

The rapid spread of fake and misleading reviews on the internet seriously endangers consumers' trust as well as the integrity of e-commerce systems. Hence, there is an urgent need for powerful, scalable, and privacy-respecting methods for detecting fake reviews. Most of the existing methods focus on either analyzing the text or the behavior of the reviewers only. They completely overlook the intricate network of interactions and the orchestrated fraudulent behaviors that are typical of sophisticated fake review campaigns. This paper introduces FedGraphFake, a federated graph-based multi-modal system for detecting fake reviews. It combines behavioral, linguistic, and relational data for fake review detection in a privacy-preserving manner. The rapid spread of fake and misleading reviews on the internet seriously endangers consumers' trust as well as the integrity of e-commerce systems. Hence, there is an urgent need for powerful, scalable, and privacy-respecting methods for detecting fake reviews. Most of the existing methods focus on either analyzing the text or the behavior of the reviewers only. They completely overlook the intricate network of interactions and the orchestrated fraudulent behaviors that are typical of sophisticated fake review campaigns. This paper introduces FedGraphFake, a federated graph-based multi-modal system for detecting fake reviews. It combines behavioral, linguistic, and relational data for fake review detection in a privacy-preserving manner. Our proposed framework creates a heterogeneous interaction graph consisting of user, product, and review nodes, which helps in spotting coordinated review patterns by using Graph Attention Network (GAT) embeddings. At the same time, the semantic representations of the review content are obtained through the fine-tuned RoBERTa encoders, and the user behavioral features such as the number of reviews published, temporal burstiness, and rating deviation are utilized to make the detection more reliable. These heterogeneous feature representations are combined through the cross-modal attention fusion mechanism and fed to a hybrid deep learning classifier that includes graph convolutional, transformer, and fully connected layers.

Federated learning (FL) is a special technique that allows client devices to collaboratively train a central model without sharing any private or raw data. To protect data confidentiality and achieve system scaling efficiently, FedAvg along with ( , )differential privacy guarantees is used as a federated learning framework, which makes it possible for several clients to jointly train the shared global model without exchanging any original user data. Extensive experiments on standard datasets such as Yelp Chi and Amazon review corpora confirm that the proposed FedGraphFake framework outperforms top centralized and single-modal baselines, achieving an F1-score of X% and AUC-ROC of Y% on the Yelp Chi dataset, respectively. Based on explainability analysis utilizing SHAP, it has been found that graph-structural features yield the most significant predictive signals, while semantic and behavioral modalities supply the next highest signals, which allows for providing detection decisions that can be interpreted and audited. Our findings highlight the power of multi-modal federated graph learning as a scalable and reliable method for detecting fake reviews in real-world e-commerce scenarios.