DeepSquid is a custom deepfake detection RNN designed to help users identify fake YouTube videos. Integrated with a Gradio web demo, DeepSquid allows users to simply paste in a YouTube URL to detect potential deepfakes. We saw promising results during testing, as the RNN model surpassed the Deepstar example models in training and validation accuracy. The model also reached 100% validation accuracy during training.
Links:
Motivations:
I was enthusiastic to do this project because the popularization of generative AI has made separating what is real and what is fake online much more difficult. I think this confusion can have serious real world implications, and I wanted to gain some experience in detecting AI-generated artifacts. I also wanted to try using TensorFlow and Keras, all my previous neural network projects have been in PyTorch.
Implementation:
The model was implemented in Python using Keras, TensorFlow, and OpenCV. Training was done on the Deepstar dataset of real and deepfaked YouTube videos.
Future Work:
- Implement a "false response" flagging system to start building a dataset of videos which the model currently fails to classify
- Retrain the model on a larger dataset, including videos created by more advanced deepfake models