This post is also available in: עברית (Hebrew)
Deepfake refers to artificial intelligence-synthesized, hyper-realistic video content that falsely depicts individuals saying or doing something. Most state-of-the-art deepfake video detection and media forensics methods are based upon deep learning, which have many inherent weaknesses in terms of robustness, scalability and portability.
US Army researchers developed a Deepfake detection method that will allow for the creation of state-of-the-art Soldier technology to support mission-essential tasks such as adversarial threat detection and recognition.
This work specifically focuses on a lightweight, low training complexity and high-performance face biometrics technique that meets the size, weight and power requirements of devices Soldiers will need in combat.
The innovative technological solution called DefakeHop was developed by researchers at the U.S. Army Combat Capabilities Development Command, known as DEVCOM, Army Research Laboratory, in collaboration with Professor C.-C. Jay Kuo’s research group at the University of Southern California.
According to ARL researcher Dr. Suya You, “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”. “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.”
Combining team member experience with machine learning, signal analysis and computer vision, the researchers developed an innovative theory and mathematical framework, the Successive Subspace Learning, or SSL, as an innovative neural network architecture. SSL is the key innovation of DefakeHop, researchers said.
Most current state-of-the-art techniques for deepfake video detection and media forensics methods are based on the deep learning mechanism, DefakeHop is built upon the entirely new SSL signal representation and transform theory. It is mathematically transparent since its internal modules and processing are explainable.
“We expect future Soldiers to carry intelligent yet extremely low size-weight-power vision-based devices on the battlefield,” You said. “Today’s machine learning solution is too sensitive to a specific data environment. When data are acquired in a different setting, the network needs to be re-trained, which is difficult to conduct in an embedded system. 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,” as reported by eurekalert.org.