With the increasing prevalence of AI-generated media, the need for robust detection mechanisms is paramount. This project aims to tackle the challenges of identifying GAN-generated deepfake images and AI-generated audio clips. By developing custom-built Convolutional Neural Networks (CNNs) and deploying them on Streamlit, this project offers an accessible and efficient solution for detecting deepfakes.
- Preprocessing: To enhance the model's ability to learn intrinsic features, the image dataset was preprocessed with Gaussian noise and blur.
- Model: A custom-built CNN with the following layers:
- Convolutional layers with ReLU activation
- Batch Normalization
- Dropout layers for regularization
- Accuracy: The model achieved an accuracy of 94%.
- Preprocessing: 2-second audio clips were transformed into respective spectrograms to capture time-frequency representation.
- Model: Another custom-built CNN with a similar architecture to the image detector, including:
- Convolutional layers with ReLU activation
- Batch Normalization
- Dropout layers for regularization
- Accuracy: The model achieved an accuracy of 85%.
Both models have shown high accuracy in their respective tasks, demonstrating the effectiveness of the preprocessing techniques and the custom CNN architectures in detecting AI-generated media.
- Image Dataset: The dataset which we used is “140k Real and Fake Faces” from Kaggle [Image Dataset Link]
- Audio Dataset: The dataset which we used is "for-2sec.tar.gz" from York University [Audio Dataset Link]
Both detectors are deployed on Streamlit, providing an easy-to-use web interface for real-time detection.