Unmasking Fake Users on Social Media: An AI-Driven Detection Framework

The widespread use of fake user accounts on social media platforms poses serious threats to digital trust, public opinion, and platform integrity. This paper presents a robust, AI-powered detection framework that integrates behavioral analytics, contextual text representation, and social graph structure learning. Our model combines a fine-tuned BERT (Bidirectional Encoder Representations from Transformers) model for textual analysis and a Graph Neural Network (GNN) built with PyTorch Geometric to capture network-level anomalies. Behavioral metadata such as account age, follower count, and posting frequency are also incorporated into the feature set. We employ SHAP (SHapley Additive exPlanations) for explainability, allowing detailed attribution of predictions to specific input features.The framework is evaluated using two public benchmark datasets: the Cresci-2017 dataset and a manually-labeled Twitter dataset, both of which include profile metadata, textual content, and interaction histories. All experiments were conducted on an Ubuntu 22.04 workstation with an NVIDIA RTX 3090 GPU and 64GB RAM. Our hybrid BERT+GNN model achieved state-of-the-art performance, with 94% accuracy and an F1-score of 0.93, significantly outperforming Random Forest, SVM, and single-modality deep learning baselines. We further analyze fake user behavior through heatmaps and word cloud visualizations.This study provides a scalable and explainable solution for detecting fake users, with potential applications in real-time moderation, bot detection, and information credibility assessment. Future work will focus on multimodal content integration (e.g., images, videos), real-time system deployment, and adaptive learning against evolving threat behaviors.