AI discovers deepfake video of Ukrainian President Volodymyr Zelenskyy

Deepfake detectors designed to recognize unique facial expressions and gestures can spot doctored videos of world leaders like Volodymyr Zelenskyy and Vladimir Putin

go through Jeremy Xu

Deepfake detectors can identify fake videos of Ukrainian President Volodymyr Zelenskyy with high accuracy. Not only could the detection system protect Zelenskyy, who was the target of deepfake attempts in the early months of Russia’s invasion of Ukraine, but it could also be trained to flag deepfakes of other world leaders and business tycoons.

“We don’t have to separate you from a billion people – we just have to separate you from [the deepfake made by] Anyone trying to imitate you,” said Hani Farid at UC Berkeley.

farid worked with Mathias Bohacek Developed face detection capabilities at the Johannes Kepler Gymnasium in the Czech Republic, sound, gestures and upper body movements.Their research builds on previous work in which AI systems were trained to Detect deep fake faces And head campaigns from world leaders like former President Barack Obama.

Boháček and Farid trained a computer model on more than eight hours of Zelenskyy videos that were previously publicly released.

The detection system scrutinizes many 10-second clips taken from a single video, analyzing up to 780 behavioral characteristics. If it flags multiple clips in the same video as fake, that’s a signal for human analysts to watch carefully.

Based on the real videos the AI ​​was trained on, it can detect when a person’s usual habits don’t match. “[It] It’s okay to say, ‘Ah, we observed President Zelensky raise his right eyebrow when he raised his left hand, and we didn’t see that,'” Farid said. Journalists or NSA analysts, they have to be able to look at this, ‘Why does it think it’s fake? ‘”

The holistic head and upper body analysis of deepfake detectors is ideal for spotting manipulated videos, and could complement commercially available deepfake detectors that primarily focus on spotting less intuitive patterns involving pixels and other image features, says Lu Siwei At the University at Buffalo in New York, he was not involved in the study.

“So far, we haven’t seen any examples of deepfake generation algorithms that can create realistic human hands and exhibit the dexterity and gestures of real humans,” Lyu said.This gives state-of-the-art detectors an advantage in Catch up on deepfakes today This fails to convincingly capture the connection between a person’s facial expressions and other bodily movements when they speak — and likely gets ahead of the rapid advances in deepfake technology.

The deepfake detector achieved 100 percent accuracy when tested on three deepfake videos of Zelenskyy that modified his mouth movements and spoken language from Delaware-based Colossyan, a company featuring AI actors. Featured custom video. Likewise, the detector performs perfectly on actual deepfakes released in March 2022.

But the time-consuming training process, which requires recording hours of video for each person of interest, is not well suited to spotting deepfakes or videos of non-consensual sex acts involving ordinary people. “The more futuristic goal is how to make these technologies work for people who don’t have as much video data,” Boháček said.

Researchers have built yet another deepfake detector focused on spotting fake videos of U.S. President Joe Biden, and are considering offerings for Russia’s Vladimir Putin, China’s Xi Jinping and billionaire Elon Musk. and other public figures to create similar models. They plan to make the detector available to certain news organizations and governments.

Journal reference: Member of the National Academy of Sciences, DOI: 10.1073/pnas.2216035119