Playing Offense and Defense with Deepfakes

Conference:  BlackHat USA 2019



The presentation discusses the offensive and defensive use of deepfakes, including the creation of deepfakes, examples of offensive use, and methods for detecting deepfakes.
  • Deepfakes are a relatively new technology that can be used for offensive purposes, such as influencing elections
  • The presentation provides a step-by-step breakdown of how deepfakes are created
  • Examples of offensive use of deepfakes are presented, including a video of a politician being impersonated
  • Methods for detecting deepfakes are discussed, including looking at images at the signal, physical, and semantic level
  • A novel approach to detecting deepfakes is introduced
  • A tool for offensive and defensive research, called deepstar, is announced and released
In 2016, researchers released a video showing how they could control the facial expressions of well-known figures. This was one of the first examples of deepfakes. In 2018, a comedian created a video in which Obama was deepfaked, sparking concerns about the use of deepfakes in politics. In June of the same year, a deepfake video of Mark Zuckerberg was released, combining both video and fake audio components.


This presentation seeks to demonstrate how deepfaking can be leveraged for offensive and defensive purposes. Specifically, deepfake internals are presented including a fine grained, step-by-step breakdown of creation, including the details of all deep learning models used in the process. A test deepfake video is created. A malicious deepfake video is then created, in which an international politician is impersonated. This video is then temporarily circulated in the wild, as a means for measuring impact. The offense section of the presentation is followed by a defense section. This section provides an overview of contemporary techniques for detecting deepfake videos. A novel approach to detecting deepfake videos is then introduced. Finally, a tool for offensive and defensive research, known as deepstar, is announced and released at the time of the presentation.