Audio Deepfake Detection

This is a deep learning project designed to detect deepfake audio content by analyzing various audio characteristics and features. The system uses a convolutional neural network (CNN) to distinguish between authentic and synthetically generated voice recordings.

deepfake-audio-detector.mp4

Dataset

The project uses the Deep Voice Deepfake Voice Recognition dataset from Kaggle:

• Dataset Link: Deep Voice Deepfake Voice Recognition
• Contains real and synthetic voice samples
• Used for training and testing the model

Features

• Real-time audio deepfake detection
• Advanced feature extraction including:
  • Mel spectrograms
  • MFCCs (Mel-frequency cepstral coefficients)
  • Spectral rolloff
  • Zero crossing rate
• Visual analysis through spectrogram plotting
• Support for various audio formats
• Pre-trained model included (best_model_v2.h5)

Dataset Structure

The project uses a structured dataset organized as follows:

data/
├── AUDIO/
│   ├── REAL/          # Original voice recordings
│   │   └── *.wav      # Real audio samples
│   └── FAKE/          # Synthetic voice recordings
│       └── *.wav      # Generated audio samples
└── DATASET-balanced.csv   # Metadata and labels

Requirements

• Python 3.11 or higher
• TensorFlow 2.15.0
• librosa 0.10.1
• numpy 1.24.3
• pandas 2.1.4
• scikit-learn 1.3.2
• matplotlib 3.8.2
• soundfile 0.12.1

Installation

1. Clone the repository:

HTML

git clone https://github.com/aditisingh02/Momenta.git

SSH

[email protected]:aditisingh02/deepfake-audio-detection.git

2. Install required packages:

pip install -r requirements.txt

3. Download the dataset and follow the file structure

Project Structure

deepfake-audio-detection/
├── app.py                 # main application
├── train_model_v2.py      # Model training script
├── test_model.py          # Model testing script
├── requirements.txt       # Python dependencies
├── runtime.txt            # Python dependencies
├── .streamlit/
│   └── config.toml       # theme and app settings
├── deepfake_audio_detector_v2.h5
└── best_model_v2.h5      # final model

Training the Model

To train a new model on your dataset:

python train_model_v2.py

Run the Application

python -m streamlit run app.py

Usage

Testing Audio Files

To test an audio file for authenticity:

python test_model.py --model best_model_v2.h5 --audio "path/to/audio/file.wav"

Example:

python test_model.py --model best_model_v2.h5 --audio "data/AUDIO/REAL/biden-original.wav"

Running the application

1. Launch the application using python -m streamlit run app.py
2. Upload a WAV audio file using the file uploader
3. Wait for the analysis to complete
4. View the detection results, confidence score, and audio visualizations

The training script includes:

• Data augmentation for better generalization
• Early stopping to prevent overfitting
• Model checkpointing to save the best version
• Learning rate reduction on plateau

Model Architecture

The CNN model architecture includes:

• 3 convolutional layers (32, 64, 128 filters)
• Batch normalization
• Max pooling layers
• Dropout for regularization
• Dense layers for classification

Performance

The model achieves:

• Test Accuracy: ~90.91% (based on validation results)
• AUC Score: High discriminative ability between real and fake audio
• Fast inference time (typically under 1 second per audio sample)
• Robust feature extraction with multiple audio characteristics:
  • Mel spectrograms
  • MFCCs
  • Spectral rolloff
  • Zero crossing rate

Acknowledgments

• Voice samples from Deep Voice Deepfake Voice Recognition