A Statistical Framework for Deepfake Detection

• Activity Number: 107 - SPEED: Statistical Methods, Computing, and Applications Part 1
• Type: Contributed
• Date/Time: Monday, August 8, 2022 : 8:30 AM to 10:20 PM
• Sponsor: Section on Statistics in Defense and National Security
• Author(s): Shannon Gallagher* and Catherine Bernaciak and Jeffrey Mellon and Dominic Ross
• Companies: Software Engineering Institute, Carnegie Mellon University and Software Engineering Institute, Carnegie Mellon University and Software Engineering Institute, Carnegie Mellon University and Software Engineering Institute, Carnegie Mellon University
• Keywords: deepfakes; image processing; neural nets; classification

Abstract

Believable media created by deep neural networks, also known as deepfakes, have been increasingly studied in the past five years. Deepfakes are a threat to national security both at the leadership level (e.g. puppeteering a state leader) and also the citizen level (e.g. abusing the image and likeness of a person in explicit media). Data-driven deepfake detection is, then, an important step in neutralizing the threat of deepfakes, especially as billions of images and over half a million hours of video are updated daily. To streamline and standardize the deepfake detection process, we establish a statistical framework using existing deepfake detection methods which account for the computational complexity of algorithms; their accuracy, precision and recall; and the features used as input. We demonstrate the use of our framework on open-source data sets including Stylegan2 images and Deepfake Detection Challenge videos and show a number of intra-dataset results. We describe how our results can be reproduced using best practices and how are framework can be simply transferred to other agencies.