ViT-Xplain: A Transparent Deepfake Detector for Consumer Electronics Based on Attention and Explainable AI

Consumer electronics platforms like social media, mobile apps and smart devices are becoming more vulnerable to deepfake videos, which can seriously harm content authenticity, user trust and legal safety. Many existing deepfake detection systems use deep neural networks that only give a final score, without explaining how the decision was made. This makes it difficult to use these systems in the real world, where people need clear reasons to trust what the model says. Problems like shortcut learning, video compression issues and background distractions also make these models less reliable. While some researchers have used convolutional and hybrid models to improve detection accuracy, they often ignore the need for explainable results. In this work, we introduce an explainable Vision Transformer (ViT) system that combines strong detection ability with clear, easy-to-understand explanations. The framework consists of five phases: dataset analysis and frame sampling, model training, evaluation and error analysis, attention roll-out and aggregation and explainability and trust scoring. Alongside predictions, the system produces attention-based visual overlays and quantitative metrics such as attention-to-face overlap, entropy concentration, clarity index and a trust score. Evaluated on the FaceForensics++ benchmark, our ViT model achieves 74.4% accuracy, 66.0% precision, 90.7% recall, 76.4% F1-score, 81.8 ROC AUC and 72.3 PR AUC. Attention maps consistently highlight facial regions with an average overlap score of 0.575 and higher alignment correlates with classifier confidence. The framework is lightweight, Python-based and deployable on standard hardware, making it practical for consumer platforms that require explainability, trust calibration and compliance-friendly audit features.