Classifying a video whether its deepfake or not
As the use of Gpu's increased for heavy mathematical computation involved in the training of deep learning models, these models have witnessed tremendous growth for scaling and making more complex models. And since everyone has access to these high computational Gpu's, almost anyone can start learning and making deep learning models. one such application
is the creation of deepfakes which means that we use deep learning techniques to manipulate and alter the expressions and facial features of people real videos with such an accuracy thet it sometimes becomes almost impossbile to tell the difference between real and fake videos through naked eye.
We propose that since AI has created this problem, perhaps we should let AI only to solve this problem for us. We have trained an EfficientNet on face-forensics++ dataset with about 1000 real and 1000 fake videos. Our model reaches an accuracy of ~90%. We furthur noticed that some videos being bad in quality, our model was not able to generalise properly, hence we took the quality of video also into account. During data preprocessing, we observed that the video blinking rate of people in real videos and fake videos differ, although not by a great margin, but still it seemed a good variable and let our logistic model worry about the variable relations.