Researchers have developed an AI algorithm that can spot deepfake videos with 98% accuracy.
In a paper published earlier this month, researchers from the Multimedia and Information Security Laboratory (MISL) at Drexel University's College of Engineering explained that they had developed the “MISLnet algorithm” that can detect telltale signs of deepfakes and manipulated media with incredible accuracy.
The team trained machine learning algorithms to extract and recognize the digital “fingerprints” of various video generators, including Stable Video Diffusion, Video-Crafter, and Cog-Video.
Moreover, the researchers showed that the algorithm can learn to detect new AI generators by studying just a few examples of created videos.
The difficulty of detecting deepfakes
according to Live ScienceThe MISLnet algorithm marks an important new milestone in detecting fake image and video content because many of the “digital breadcrumbs” that existing systems look for in regular digitally edited media do not exist in fully AI-generated media.
“When an image is created, the physics and algorithmic processing of the camera introduces relationships between different pixel values that are very different from the pixel values you'd see if you generated an image with Photoshop or AI,” Dr. Matthew Stam, associate professor in Drexel University's College of Engineering and director of MISL, said in a statement.
“But more recently, text-to-video tools like Sora have emerged that allow you to create some pretty impressive videos. And because they're not made in camera or Photoshopped, they pose a whole new challenge.”
“Until now, forensic detection programs have worked well on edited videos by simply treating it as a series of images and applying the same detection process,” Stam adds.
“But because AI-generated videos lack frame-by-frame evidence of image manipulation, to be effective detection programs must be able to identify new traces left by the way generative AI programs construct videos.”
Because AI-generated videos are not produced by a camera capturing real scenes or images, they do not contain noticeable variations between pixel values.
but, Live Science The team's new MISLnet algorithm was reportedly trained 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.
The MISL algorithm was able to correctly detect the AI-generated videos 98.3% of the time, outperforming eight other systems created by the research team that scored above 93%.
The research team has been active in flagging digitally altered images and videos for more than a decade, but has been particularly busy in the last year as editing techniques have been used to spread political misinformation.
“I'm a little worried [AI-generated video] “Counterfeits created by bad actors could be released before we have good systems in place to detect them,” Stam said.
