CLFT: An Optimized Hybrid Cross-Layer Fusion Transformer for Accurate Fake Profile Detection on Social Media

Abstract

The rapid increase of fake profiles on social media platforms has raised significant concerns regarding online authenticity, user trust, and digital security. Despite various efforts to combat this issue, existing detection methods often fall short due to the evolving nature of fake profiles and the noisy, high-dimensional data involved. This study proposes an optimized Hybrid Cross-Layer Fusion Transformer (CLFT) for detecting fake profiles by analyzing behavioral metadata. The CLFT architecture integrates multi-stage attention mechanisms, including Cross-Layer Fusion Attention (CLFA), Sparse–Dense Hybrid Attention (SDHA), and Temporal-Behavior Embedding Blocks (TBEB), to effectively capture both short- and long-term dependencies in user activities. The model hyperparameters were optimized using the Bayesian Optimization Hyperband (BOHB) framework. Experimental results on a real-world social media dataset show that the proposed model outperforms traditional machine learning techniques and previous Transformer-based models, achieving an accuracy of 99.10%, precision of 99.89%, recall of 99.55%, and an F1-score of 99.72%. Furthermore, the attention mechanisms enhance interpretability by emphasizing the most influential behavioral features, contributing to the model’s transparency and reliability. The findings highlight that Transformer-based models, especially the CLFT, provide a scalable and efficient solution for fake profile detection in noisy environments, with important implications for enhancing social media security. The study emphasizes the need for interpretability in automated detection systems, fostering trust and ensuring better user engagement and platform integrity.