Overview This project focuses on detecting deepfakes using a comprehensive approach that combines video, audio, and image data. While many existing solutions focus on a single method, our model integrates all three to enhance detection accuracy and robustness. The performance metrics achieved are impressive, with video detection performing well, image detection achieving 90.85% accuracy, and audio detection reaching 99.05% accuracy.
Features Multimodal Detection: Integrates video, audio, and image data for deepfake detection.
Comprehensive Analysis: Utilizes multiple data types to improve the reliability of deepfake detection.
Prerequisites To run this project, you will need:
Other necessary libraries specified in requirements.txt
bash Copy code
cd deepfake-detection-multimodal
Install the required libraries:
bash
Copy code
pip install -r requirements.txt
Usage
Data Preparation: Ensure you have your video, audio, and image data organized. Follow the instructions in data_preparation.md for guidance on how to prepare your datasets.
Training: To train the model, run:
bash
Copy code
python train.py --data_path /path/to/your/data
Evaluation: To evaluate the model on your test data, use:
bash
Copy code
python evaluate.py --model_path /path/to/your/model --data_path /path/to/your/test/data
Prediction: For predicting whether a given input is a deepfake, run:
bash
Copy code
python predict.py --input_path /path/to/your/input --model_path /path/to/your/model
Results
Video Detection: Our video detection model has shown high performance, making it reliable for identifying deepfakes in video content.
Image Detection: Achieved an accuracy of 90.85%, proving its effectiveness in detecting fake images.
Audio Detection: Reached an impressive accuracy of 99.05%, highlighting its strength in identifying manipulated audio.
Model Architecture
The project utilizes a combination of convolutional neural networks (CNNs) for image and video data and recurrent neural networks (RNNs) for audio data. The models are fine-tuned using transfer learning to leverage pre-trained weights and improve accuracy.
We welcome contributions to enhance the project. Please fork the repository and create a pull request with your changes. Ensure that your contributions are well-documented and include relevant test cases.
This project is licensed under the MIT License. See the LICENSE file for details.
We would like to thank the open-source community and the creators of the pre-trained models used in this project.
For any questions or feedback, please contact Debojyoti Bhuinya, Subhamay Ganguly, Akash Das, Gargi Chakraborty.