A Multimodal Framework for Explainable Fake-Profile Detection in Online Social Networks

Abstract

Fake profiles and social bots increasingly combine plausible text, convincing images, regular-looking activity and carefully chosen social connections. A detector that examines only one source of evidence can therefore miss accounts that are deliberately designed to look ordinary. This paper presents a multimodal framework that combines four complementary evidence streams: semantic content, visual information, temporal behaviour, and social-graph structure. Transformer-based text representations, visual encoders, temporal activity features and graph neural representations are projected into a shared latent space. A profile-specific attention mechanism then assigns weights to the available modalities, while an explicit availability mask distinguishes missing evidence from genuinely low activity. The framework is designed for explainable decision support rather than an opaque binary verdict; modality contributions, feature attributions and neighbourhood evidence are intended to support analyst review. The supplied manuscript reports aggregate accuracy values of 85.2% for SVM, 88.6% for Random Forest, 91.4% for CNN, 93.8% for GNN and 96.7% for the proposed multimodal framework. These values are retained as reported evidence and are not presented here as independently reproduced experiments. The principal contribution of this manuscript is the integrated architecture, the missing-modality design, and a publication-oriented validation protocol that explicitly addresses ablation, calibration, temporal drift, cross-platform transfer, adversarial manipulation, and reproducibility.