New biometric tool spots deepfakes even out on the battlefield

A new method of foiling deepfake video reportedly has been developed by Army researchers.

The system, called DefakeHop, is billed as being lightweight, requiring little training and offering high performance for operating in combat. Army personnel are increasingly put on the ground with sophisticated vision systems that could fall victim to digital attacks.

DefakeHop is a unique tool for processing and understanding face biometrics.

The Army Research Lab collaborated with University of Southern California scientists on the project. Army researcher Suya You was looking for a new way to understand what makes deepfakes so realistic, and, of course to find defenses against them.

The scientists developed “an innovative theory and mathematical framework” called the Successive Subspace Learning (SSL) neural net architecture. SSL architecture consists of multiple transform matrices processed in cascade. The method is expected to overcome the shortcomings of deep-learning media forensics, including scalability, portability and robustness (in terms of adversarial attacks).

The team also developed the FaceHop algorithm based on SSL to improve the performance of biometric gender classification with low image quality.

Researchers will present the method in July at the IEEE International Conference on Multimedia and Expo 2021.

Article Topics

biometrics  |  biometrics research  |  deepfakes  |  facial recognition  |  spoof detection  |  U.S. Army