Deepfake Detection

A deep learning-based project for detecting deepfake videos and images, focusing on robustness under social media compression. This project evaluates multiple state-of-the-art CNN models for deepfake detection and provides tools for model training and evaluation.

Features

• Multiple model architectures: XceptionNet, EfficientNet (B4, B7), MesoNet, and ResNet
• Support for different datasets including DeepfakeTIMIT and WildDeepfake
• Pre-processing utilities for video frame extraction and data balancing
• Model evaluation and comparison tools
• Pre-trained models for quick deployment

Datasets

DeepfakeTIMIT (LQ)

• Source: IDIAP Research Institute
• Description: Low-quality (64×64) deepfake videos
• Content: Manipulated videos based on VidTIMIT dataset

VidTIMIT

• Source: VidTIMIT
• Description: Source dataset with real videos
• Content: Videos of 43 consented volunteers

Installation

1. Clone the repository:

git clone https://github.com/yourusername/Deepfake_Detection.git
cd Deepfake_Detection

2. Create and activate a virtual environment (recommended):

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install the required packages:

pip install -r requirements.txt

Project Structure

Deepfake_Detection/
├── VidTIMIT/               # Code for models trained on VidTIMIT dataset
│   ├── ResNet152.py        # ResNet model implementation
│   ├── efficientNetB4_model.py  # EfficientNetB4 implementation
│   ├── efficientNetB7.py   # EfficientNetB7 implementation
│   ├── mesoNet.py          # MesoNet implementation
│   └── xception_model.py   # Xception model implementation
├── VidTIMIT_deepfake_dataset/  # DeepfakeTIMIT dataset
├── WDFtest/                # WildDeepfake test implementations
├── WildDeepfake/           # Code for models trained on WildDeepfake dataset
├── weights/                # Trained model weights
│   ├──  VidTIMIT
│   └──  WD
├── pre_processing/         # Data preprocessing utilities
│   ├── balance_frames.py   # Utility for balancing dataset
│   └── extract_frame.py    # Frame extraction from videos
├── comparison.py           # Model comparison utilities
└── kaggle.py               # WildDeepfake competition utilities
└── charts/                 # Charts for model comparison

Usage

Data Preparation

1. Extract Frames from Videos:

python pre_processing/extract_frame.py --input_dir /path/to/videos --output_dir /path/to/save/frames

2. Balance Dataset:

python pre_processing/balance_frames.py --data_dir /path/to/frames --output_dir /path/to/balanced_data

Training a Model

Example: Training Xception model on VidTIMIT dataset

python VidTIMIT/xception_model.py --data_dir /path/to/training_data --model_save_path /path/to/save/model

Evaluating a Model

Example: Evaluating a pre-trained model

python VidTIMIT/xception_test.py --model_path /path/to/model --test_data /path/to/test_data

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Acknowledgements

• DeepfakeTIMIT dataset provided by IDIAP Research Institute
• VidTIMIT dataset