The DefakeHop tool combines machine learning, signal analysis and computer vision.
The US Army has introduced a lightweight deepfake detection tool to prevent potential threats to national security.
“With the advancement of generative neural networks, AI-driven deepfakes are evolving so rapidly that there are simply not enough reliable methods to detect and defend against them. There is an urgent need for an alternative paradigm that can understand the mechanism behind the astounding performance of deepfakes and develop effective security solutions with strong theoretical support, ”explained Chung-Chi professor of electrical engineering and computer science at the University of Southern California Chung-Chi Jay Kuo (C.-C. Jay Kuo). …
A team of experts led by Kuo created DefakeHop, a tool that combines machine learning, signal analysis and computer vision. DefakeHop is based on the new Successive Subspace Learning (SSL) neural network architecture. DefakeHop’s unique design has several advantages over traditional deepfake detection methods, including greater transparency, lower control needs, smaller model sizes, and better security, the researchers said.
As Kuo explained, SSL represents a completely new mathematical framework for neural network architecture, developed based on signal transformation theory.
“It is radically different from the traditional approach, offering a new signal representation and process that includes multiple transformation matrices in a cascade. […] It is a complete, uncontrolled data-driven environment that offers a completely new tool for image processing and understanding tasks such as facial biometrics, ”said Kuo.
In addition to detecting deepfakes, experts have found a number of uses for their proposed lightweight model of interpreting images in the military.
“We expect that in the future, soldiers will use intelligent but extremely lightweight and powerful vision devices on the battlefield. The proposed solution has a number of desirable characteristics, including a small model size, limited training data, low training complexity, and the ability to process input images at low resolution. This could lead to the development of revolutionary solutions with far-reaching applications for the army, ”explained researcher at the Army Research Laboratory (ARL) Suya Yu.