Army Claims Breakthrough in Deepfake Detection
The Army has developed a new deepfake detection technology designed to support soldiers in tasks such as adversarial threat detection and recognition.
With help of researchers from the University of Southern California, the Army Research Laboratory developed a lightweight, low-complexity face biometrics tool built for soldiers in combat, the Army said.
The DefakeHop tool operates using a mathematical framework called successive subspace learning, which a lead researcher described as being “radically different from the traditional approach.”
C.-C Jay Kuo, a distinguished professor at USC, said SSL is a data-driven unsupervised framework that serves as a cutting-edge method for image processing and face biometrics.
ARL researcher Suya You said that a novel approach is needed to keep up with advancements in generative adversarial networks and artificial intelligence-driven deepfake technologies.
According to You, even the best anti-deepfake solutions today have weaknesses in robustness, scalability and portability due to their reliance on deep learning.
You added that most current machine learning solutions are only effective in specific environments and would be too difficult to constantly retrain in embedded solutions such as DefakeHop.
The researchers claimed that SSL is superior in that it offers “mathematical transparency,” lower training complexity, smaller model sizes and parameters, and better defenses against adversarial attacks.
The technology was featured in a paper titled “DefakeHop: A Light-weight High-performance Deepfake Detector,” which will be presented at the IEEE International Conference on Multimedia and Expo 2021.
According to the Army, the ARL-USC research team will continue pursuing breakthroughs in face biometrics, target detection, recognition and semantic scene understanding.
Category: Defense and Intelligence
Tags: Army Army Research Laboratory artificial intelligence C.-C Kay Kuo deep learning deepfake DefakeHop Defense and Intelligence machine learning SSL successive subspace learning Suya You University of Southern California