A Unified Deep Learning Model for Fake Account Identification Using Transformer-Based NLP and Graph Neural Networks

Fake accounts and bots on social media platforms pose significant threats, including misinformation dissemination, fraud, and manipulation of public opinion. To address these challenges, a hybrid deep learning model integrating Transformers and Graph Neural Networks (GNNs) is proposed. The model leverages BERT for textual analysis of user-generated content, including posts, comments, and profile descriptions, while GNNs analyze network-based features such as follower relationships, engagement behaviors, and interaction patterns. This combination ensures an enhanced understanding of both linguistic and structural characteristics, leading to improved detection accuracy. The dataset is curated from publicly available labeled social media account data, supplemented with API-based extractions to ensure diversity. Features include user metadata, textual content, and network structures. In the preprocessing phase, text is tokenized using a BERT tokenizer and converted into high-dimensional embeddings using BERT-based encoders. Simultaneously, a social graph is constructed, where users and interactions are represented as nodes and edges. These representations serve as input for the GNN model, learning patterns indicative of fraudulent activities. A dual-module architecture is implemented, comprising a BERT-based text classification model and a GNN-based network classification model. Outputs from both modules are fused in a shared feature space before classification. Model performance is evaluated using accuracy, F1-score, and AUC-ROC metrics. Experimental results demonstrate superior detection capabilities compared to singlemodality models. Applications include social media security, fraud prevention, misinformation control, and e-commerce integrity. Future enhancements focus on real-time scalability and robustness against adversarial attacks.