🎭 Deepfake Video Detection MVP (Self-Trained ViT)

This repository contains a self-trained deepfake video detection system built using a Vision Transformer (ViT-B/14 with DINOv2 backbone).
The system analyzes videos temporally, identifies manipulated segments, and outputs timestamp-localized deepfake regions, with an emphasis on localization quality over raw classification accuracy.

---

🚀 Key Features

• ✅ Self-trained deepfake model (not a prebuilt classifier)
• 🎯 Vision Transformer (ViT-B/14, DINOv2)
• ⏱️ Timestamp localization of manipulated segments
• 📊 Median smoothing + temporal segment merging
• 🎞️ Video-level and segment-level confidence scores
• 🖥️ Interactive Streamlit web interface
• ⚡ Efficient inference via 2 FPS frame sampling

---

🧠 Model Overview

Component	Description
Backbone	vit_base_patch14_dinov2
Framework	PyTorch + TIMM
Input Resolution	518 × 518
Output	Binary classification (Real / Fake)
Weights	Self-trained (df_detector_mvp.pth)

Each frame produces a single logit, converted to a probability using a sigmoid function.