A full-fledged system for detecting deepfakes in images and videos using a custom-trained deep learning model powered by EfficientNetB0, wrapped in an interactive Streamlit interface.

This project implements a deepfake detection pipeline from scratch. It includes:

• A training pipeline using EfficientNetB0 with fine-tuning.
• Real-time image and video classification using a Streamlit web app.
• Frame-by-frame video analysis for deepfake frame percentage.
• Robust handling of grayscale, RGB, and RGBA inputs.

The goal is to showcase applied machine learning, model deployment, and frontend integration for a practical AI application.

├── train/ # Model architecture, training scripts
├── test data/ # Sample test videos and images
├── streamlit/app.py # Streamlit frontend code for inference
├── streamlit/my_model.keras # Saved trained model

1. Training Phase

   • EfficientNetB0 (imagenet weights) is used as a base model.
   • Custom dense layers are added for binary classification.
   • Trained on labeled real vs fake datasets.

2. Streamlit App

   • Users upload an image or video.
   • For images: Model instantly classifies as Fake or Real.
   • For videos: Each frame is predicted, and a percentage of fake frames is shown.

• Python
• TensorFlow / Keras
• Streamlit
• OpenCV
• NumPy / PIL

⚠️ This project is built for educational purposes and is not intended for production-grade deepfake detection.