With the release of artificial intelligence Thanks to (AI) video generation products like Sora and Luma, we are on the verge of a deluge of AI-generated video content, and policymakers, public figures, and software engineers are already caveat Deepfakes are rampant. With algorithms identifying telltale signs of AI videos with over 98% accuracy, it seems the best defense against AI counterfeiting may be AI itself.
The irony of protecting us from AI-generated content is hard to miss, says project leader Matthew Stam, an associate professor of engineering at Drexel University. statement“That's a bit unsettling. [AI-generated video] Counterfeit products may be released before we have good systems in place to detect the counterfeits created by bad actors.”
“Until now, forensic detection programs have been effective against edited videos by simply treating them as a series of images and applying the same detection process,” Stam added, “But with AI-generated videos, there's no frame-by-frame evidence of image manipulation, so for a detection program to be effective it needs to be able to identify new traces left by the way the generative AI program constructs the video.”
The breakthrough was outlined in a study published April 24. Preprint server arXivThe algorithm marks an important new milestone in detecting fake image and video content, as many of the “digital breadcrumbs” that existing systems look for in regular digitally edited media don't exist in fully AI-generated media.
Related: 32 times artificial intelligence has made catastrophic mistakes
The research project's new tool for deepfakes, called MISLnet, evolves from years of data gained from detecting fake images and videos using tools that detect changes made to digital videos and images, which can include adding or moving pixels between frames, manipulating the speed of clips, or removing frames.
Such tools work because algorithmic processing in digital cameras creates relationships between pixel color values that can be very different in user-generated images or images edited in apps like Photoshop.
But because AI-generated videos are not generated by a camera capturing real scenes or images, they won't contain noticeable differences between pixel values.
The Drexel team's tools, including MISLnet, train using a method called constrained neural networks, which can distinguish between normal and abnormal values at the sub-pixel level in images and video clips, rather than looking for general indicators of image manipulation like those described above.
MISL outperformed seven other fake AI video detection systems, correctly identifying AI-generated videos 98.3% of the time, beating eight other systems that scored at least 93%.
“We are already seeing AI-generated videos being used to create disinformation,” Stam said in a statement. “As these programs become more widespread and easier to use, we can reasonably expect to see an increase in synthetic videos.”
