Building a tool to spot Deepfakes

Maintaining the nation’s security continues to be a concern of high priority in the United States. This is particularly true in cyberspace, where information is constantly changing and being manipulated. One example of this is deepfake technology, where the facial features of important figures are changed, typically to sway public opinions and influence decision making.

Dr. Shuyan Metcalfe (publications under Shuyan Mary Ho), an associate professor at FSU’s College of Communication and Information, and her associates have been tackling issues of trusted human-computer interaction. Their research focuses on computer-mediated deception, through examining physical and language components. Her most recent project, funded by the U.S. Air Force, focuses on addressing the problem of deepfake image alteration at the pixel level using artificial intelligence and machine learning.

Understanding and detecting image manipulation is a complex and under-researched problem. Originally, this problem was in the domain of forensics professionals, but the high computing power required has shifted it to Dr. Metcalfe’s area of expertise. Through examining image manipulation at the pixel level, the work done by Dr. Metcalfe and her team is a trailblazing effort in the intelligence analysis community.

Dr. Metcalfe and her team create, train and run machine-learning models for this project through the FSU High Performance Computing (HPC) cluster with the assistance of the Research Computing Center (RCC) staff. Using the PyTorch software package, Dr. Metcalfe’s team can run their models and detect when images are manipulated. Their ultimate goal is to develop a user experience (UX) infrastructure that allows users to upload images through the web and mobile devices and run the images through their models in the HPC environment. This will enable users to detect if an image has been manipulated in a short period of time. By researching and developing this infrastructure, Dr. Metcalfe can promote national trust and security.

“The more I study this deception problem, the more I think it is critical to afford everyone the ability to safeguard against social manipulation and seek the truth.”