Deepfake Detection

A deep learning framework for audio deepfake detection using models including AASIST and LCNN architectures.

Features

• Support for multiple model architectures (AASIST, LCNN)
• Multi-GPU and single-GPU training support
• Comprehensive evaluation metrics
• Audio preprocessing and augmentation
• Pre-trained model checkpoints

Environment Setup

Prerequisites

• Python 3.10+
• CUDA-compatible GPU (recommended)
• Conda package manager

Installation

1. Create and activate conda environment:

conda create -n deepfake python=3.10
conda activate deepfake

2. Install dependencies:

pip install -r requirements.txt

Alternatively, you can use the provided conda environment file:

conda env create -f environment.yml
conda activate deepfake

Dataset Preparation

Data Format

Prepare your dataset with the following structure:

data/
├── train.txt
├── test.txt
└── audio_files/
    ├── file1.wav
    ├── file2.wav
    └── ...

Label File Format

Create label files (train.txt, test.txt) with the following format:

file_id,file_path,label
0069156,aigc_speech_detection_tasks_part1/0069156.wav,Bonafide
0069157,aigc_speech_detection_tasks_part1/0069157.wav,Spoof

Where:

• file_id: Unique identifier for the audio file
• file_path: Relative path to the audio file
• label: Either "Bonafide" (real) or "Spoof" (fake)

Training

Multi-GPU Training

For training with multiple GPUs:

bash train.sh

Single-GPU Training

For training with a single GPU:

bash run_single.sh

Custom Training

You can also run training directly with custom parameters:

python train.py --config_path configs/your_config.yaml

Inference

Batch Inference

Run inference on a dataset:

python inference.py --model_path path/to/model.pth --data_path path/to/test_data

Single File Inference

Use the provided shell script for quick inference:

bash infer.sh path/to/audio_file.wav

Model Architecture

This project supports multiple architectures:

• AASIST: Audio Anti-Spoofing using Integrated Spectro-Temporal graph attention
• LCNN: Light Convolutional Neural Network

Project Structure

├── module/                 # Model architectures
│   ├── AASIST.py          # AASIST model implementation
│   └── LCNN.py            # LCNN model implementation
├── scripts/               # Utility scripts
├── ckpts/                 # Pre-trained checkpoints
├── exports/               # Training outputs
├── dataset.py             # Dataset loading utilities
├── model.py               # Model wrapper
├── learner.py             # Training logic
├── inference.py           # Inference utilities
├── metrics.py             # Evaluation metrics
└── params.py              # Configuration parameters

Results

Training logs and model checkpoints are saved in the exports/ directory. You can monitor training progress and evaluate model performance using the provided metrics.

Contributing

Feel free to submit issues and pull requests to improve this project.

License

This project is licensed under the MIT License - see the LICENSE file for details.