In this repository, I have worked on deepfake image detection using MTCNN and InceptionResnetV2.
The links to the weights file and the google colab code has been provided in this repository.
Here, a frame extraction was performed on the training data from the Deepfake Detection Challenge from Kaggle.
Then MTCNN was used for face extraction and each cropped frame was saved as an image based on the label.
Following this, I built the InceptionResnetV2 network and used the pre-trained Imagenet weights.
This model gave me a 99.73% training accuracy and a 93.58% validation accuracy.
The loss and accuracy graphs are given below:
The output for every frame is a 2d vector which indicates how close each frame is to being a real or fake image.
This can help us decide whether the video as a whole is a fake or real.
Some future areas to work on regarding this would be to include an audio file classification as well.
We could also look into ways to increase the accuracies along with displaying all the frames in the video, along with the audio and video analysis.