Towards detecting state-of-the-art deepfakes

Deep learning, while able to solve complex problems, is likewise capable of creating technologies that threaten privacy, democracy, and national security. One of these technologies is the ability to create images or videos that humans cannot distinguish from authentic media. Such generated media are termed ‘deepfakes’. While methods of automatic deepfake detection exist, current methods are unsuitable for deployment at scale, partly due to their inordinate computational cost. This paper performs a comprehensive analysis of recent deepfake detection methods and proposes multiple significant improvements, including a novel face detector. Together these culminate in an order of magnitude improvement of efficiency over the state-of-the-art.

Category: COMPUTING Country: IRELAND Year: 2021


Gregory Guy Tarr