Scientists Use AI to Identify AI Deepfake Videos
This post is also available in: 
עברית (Hebrew)
While content generated by artificial intelligence has been prevalent online for some time now, AI video generation has recently advanced to a whole new level.
OpenAI’s Sora, which was released back in February, has been wowing users with hyper-realistic synthetic content generated only using text prompts. However, its highly convincing output makes many people “fall” for its generated content, causing rising concerns regarding the spread of misinformation.
Experts from the Multimedia and Information Security Lab (MISL) in Drexel’s College of Engineering have been working on creating technologies to manipulate imagery for over ten years. According to Cybernews, currently existing methods have been ineffective against AI-generated videos – the experts from the lab evaluated 11 synthetic image detectors that are currently available to the public and found they had at least 90% accuracy when detecting manipulated images. However, their performance decreased by 20-30% when asked to identify videos that were generated by publicly available AI generators.
PhD Matthew Stamm, director of the MISL, said in a press release: “It’s more than a bit unnerving that this video technology could be released before there’s a good system for detecting fakes created by bad actors,” adding that once the technology is publicly available, malicious usage is inevitable. “That’s why we’re working to stay ahead of them by developing the technology to identify synthetic videos from patterns and traits that are endemic to the media.”
The way MISL identified altered and AI generated media so far is by detecting the digital traces left behind after photo and video editing programs alter pixels, adjust speed, or manipulate frames. Unfortunately, the new text-to-video generators have not been produced by a camera and were not edited by visual software, and therefore pose new challenges in detecting manipulation.
In their latest research, MISL’s scientists tried using AI against AI to identify how generative AI programs construct their videos, and managed to successfully train a machine learning algorithm called “a constrained neural network” – the network was able to learn what a synthetic video looks like at a granular level, and apply that knowledge to a new set of videos that were generated with various AI video generators.
The algorithm proved effective more than 93% of the time in identifying synthetic videos and even accurately naming the program that was used to create them.
