Image Animations on Driving Videos with DeepFakes and Detecting DeepFakes Generated Animations

The concept of image animation is to create a video or animation such that an object from an image is animated as per the motion of driving video. We plan to analyze with minor modifications of an existing framework which does this without any information beforehand about the object which is to be animated. To achieve that, we train our dataset on a set of images and videos for the objects of same category, for example (face, body, street views) etc. Some recent applications of neural networks (CNN) have proved to form realistic human heads. Realistic talking heads can be created by training the dataset of large number of images and videos. A source image of a person can be animated on target poses of a person (driving video), by keeping the appearance and body of the person. However, on the parallel side, there are advancements in the development of systems which are capable of detecting DeepFakes generated videos and animations as it is a crucial security concern. We did experiments on Image Animation to achieve talking heads, Image generations with conditional generative adversarial networks for DeepFakes Generations and the results were realistic. Moreover, we implemented a DeepFake Detector XceptionNet with minor modifications which achieved 95% accuracy on detecting DeepFakes. At last, we implemented a newly introduced technique in which the DeepFake generation is perturbed through which it can easily fool the deepfake detector. XceptionNet was able to achieve less than 30% accuracy on detecting DeepFakes generations when they were perturbed.