Using Domain Knowledge RCNN for Deepfake Detection

As deep fakes and other similar synthetic image and video generation and manipulation techniques become more prominent, the world is entering a new era where media is no longer a determinant or factual evidence of trustworthiness. While media sabotage is not a new concept, the use of artificial intelligence techniques such as deepfakes makes the generation of those fake images and videos more accessible and far cheaper in terms of time and effort. This paper explores the most modern of those synthetic image and video generation technologies: deepfakes. It is vital to raise awareness about the existence of the technology, the level of accessibilities, and its impact. To limit the impact of disinformation those technologies can spread, we propose a novel Recurrent Convolutional Neural Network model that focuses on the facial and other relevant domain-specific knowledge of the video to detect deepfake videos. This novel model abstracts away from irrelevant details in the inputted video which greatly improves the consistency of the detection. Furthermore, by identifying the features deepfakes struggle in producing, the model will learn about the features and attributes to prioritize during evaluation.