U.S. Army Employs Machine Learning for Deepfake Detection
Chances are, you’ve heard about deepfakes: convincing and easy-to-manufacture facial imposition and manipulation tools that allow even moderately tech-savvy individuals to, for instance, falsify realistic video clips of politicians saying things they’ve never said. Deepfakes have concerned many communities – and in particular, governments – for some time, and now the U.S. Army is introducing a lightweight deepfake detection method to preempt the national security concerns that will arise from the technology.
“Due to the progression of generative neural networks, AI-driven deepfake advances so rapidly that there is a scarcity of reliable techniques to detect and defend against deepfakes,” explained C.-C. Jay Kuo, a professor of electrical and computer engineering at the University of Southern California. “There is an urgent need for an alternative paradigm that can understand the mechanism behind the startling performance of deepfakes and develop effective defense solutions with solid theoretical support.”
To that end, Kuo’s group created DefakeHop through a combination of machine learning, signal analysis and computer vision. DefakeHop is powered by “Successive Subspace Learning” (or SSL), a new neural network architecture that enables their deepfake detection. The researchers explained that DefakeHop’s unique construction offers it several advantages over traditional deepfake detection methods, including greater transparency, low supervision needs, smaller model sizes, and better security.
“SSL is an entirely new mathematical framework for neural network architecture developed from signal transform theory,” Kuo said. “It is radically different from the traditional approach, offering a new signal representation and process that involves multiple transform matrices in cascade. … It is a complete data-driven unsupervised framework, offers a brand new tool for image processing and understanding tasks such as face biometrics.”
Beyond deepfake detection, the researchers see a number of applications for this type of lightweight image interpretation model in the military. “We expect future Soldiers to carry intelligent yet extremely low size–weight–power vision-based devices on the battlefield,” said Suya You, a researcher at the Army Research Laboratory. “The developed solution has quite a few desired characteristics, including a small model size, requiring limited training data, with low training complexity and capable of processing low-resolution input images. This can lead to game-changing solutions with far reaching applications to the future Army.”
For now, the team is continuing its work on applications like target detection, facial recognition, and semantic scene understanding, including a new approach called FaceHop aimed at face gender classification using low-quality images. “We see this research as new, novel, timely and technically feasible today,” You said of the team’s novel approach to face biometrics. “It is a high risk, high innovation effort with transformative potential.”