The increasing realism of synthetic media has made it difficult to reliably distinguish authentic images and videos from manipulated content using manual inspection alone. This paper presents DeepFake Sentinel, an explainable Vision Transformer (ViT)-based framework for deepfake detection and attention-based evidence visualization. The proposed system supports image and short-video uploads, validates media by decoding its actual content, and performs input-quality assessment using resolution, blur, brightness, contrast, and face-availability signals. A configurable ViT-based classifier estimates the probability of manipulated content and applies threshold-based decision logic to categorize results as likely authentic, likely manipulated, or inconclusive. For interpretability, the framework employs attention rollout to generate grayscale attention maps, heatmaps, and visual overlays that indicate image regions influencing the model's prediction. These visualizations are explicitly distinguished from ground-truth forgery segmentation. For video analysis, representative frames are sampled and their predictions are aggregated using a deterministic trimmed-mean strategy, while insufficient evidence results in an inconclusive outcome. The system integrates a React and TypeScript frontend, a FastAPI backend, and PostgreSQL-based persistence to maintain analysis history, model metadata, thresholds, evidence records, reviewer decisions, and audit information. An optional large language model provides explanations based on structured detector outputs but cannot modify classification results. Prototype validation demonstrates the implementation of the system workflow and attention-evidence visualization; however, comprehensive quantitative detection performance remains subject to evaluation on licensed, diverse, labeled datasets. DeepFake Sentinel provides a traceable, human-in-the-loop framework for visual authenticity screening while explicitly representing uncertainty and the limitations of model-generated evidence.

Deepfake Detection, Vision Transformer (ViT), Deep Learning, Computer Vision, Image Forgery Detection