Detection of deepfaked images from the FaceForensics++ dataset using a pre-trained ResNet in PyTorch.
You can download the model weights from here: https://drive.google.com/uc?id=1RJMdwG1Se94hUnIiUCwgGcmgoj5x3qdH&export=download
Note from the future: Last year, I learned that the high accuracy was due to data leakage between the splits; your splits must be video-wise NOT frame-wise. Also, you must ensure a good distribution in the splits covering various scenarios to have a true test of generalizability. Thus, you will not get the expected results with the given weights :(. However, we learned something! Hope the notebook helps you out :)