Deepfake Detection System
AI-powered system that detects AI-generated fake faces in images and videos with 89.9% AUC accuracy.
Project Overview
A complete forensic-grade deepfake detection system using MobileNetV2 transfer learning. The system analyzes faces in images, videos, and batch files, showing exactly where AI manipulation artifacts exist using Grad-CAM heatmaps. Deployed live for public testing.
⚠ The Problem
With deepfake technology becoming accessible, businesses and individuals need reliable ways to verify whether media content is genuine or AI-generated.
✓ The Solution
Built a 2-phase transfer learning pipeline using MobileNetV2 trained on 3,200 real and fake face crops extracted from 400 videos. Grad-CAM heatmaps highlight the exact regions where manipulation occurs.
Results & Outcomes
Achieved Val AUC of 0.8994 — production-ready accuracy. The model correctly identifies both real and fake faces with high confidence. Live deployment allows public testing.
Visual Outputs
