Deepfake Detection System

An AI-powered platform that detects deepfake images and videos with 96%+ accuracy using EfficientNet-B4, ONNX inference, and explainable AI (Grad-CAM).

Overview

The Deepfake Detection System is a full-stack AI application designed to identify manipulated media with high accuracy and transparency. It leverages EfficientNet-B4 trained on real-world datasets (FaceForensics++, DFDC, Celeb-DF) and performs optimized inference using ONNX Runtime. The system includes image and video detection pipelines, automatic face extraction using OpenCV, and Grad-CAM explainability to visualize model decision-making. Built with FastAPI and Next.js, the platform delivers real-time predictions, detailed confidence metrics, and visual insights, making it suitable for media verification and AI-driven content moderation use cases.

Problem Statement

With the rapid rise of deepfake technology, there is a critical need for automated systems that can detect manipulated media and provide explainable insights to ensure trust in digital content.

✨ Key Features

• Detects deepfake images with 96%+ accuracy using EfficientNet-B4 trained on real-world datasets.
• Performs video deepfake detection using frame extraction and majority voting strategy.
• Generates Grad-CAM heatmaps to visually explain model predictions and highlight manipulated regions.
• Uses ONNX Runtime for fast and efficient CPU-based inference (3–4x faster than PyTorch).
• Automatically detects and crops faces using OpenCV Haar Cascade for improved accuracy.
• Provides detailed output including confidence scores, probability breakdown, and detection metadata.

Achievements

• Achieved 96.14% validation accuracy with AUC score of 0.9916 on combined deepfake datasets.
• Built end-to-end ML pipeline including data preprocessing, training, evaluation, and deployment.
• Implemented explainable AI using Grad-CAM — a feature rarely included in student projects.
• Optimized model inference using ONNX, eliminating the need for GPU in production.
• Successfully deployed backend and frontend using cloud platforms with scalable architecture.

Screenshots