The technology used to create DeepFake videos – AI-generated face swaps to make fake content – is getting better and more realistic. To combat this cycle of “fake news”, researchers from the State University of New York (SUNY) developed an AI method to tell the difference between real people saying real things and those falsely generated using smart technology.
To spot a fraud, we must first understand how DeepFake works. Training AI to make fake videos involves feeding it images rather than video. The AI then finds common ground between the two characters, stitches them together, and superimposes photos of the person being impersonated onto an actor. You might recall that one time Jordan Peele impersonated Obama (with near-perfection) and called Trump a “complete and utter dipshit”, or the series of fake celebrity porn.
Now then, how do we tell the difference? Among other things, DeepFakers don’t blink like humans.
The average human blinks 17 times per minute. Still photographs don’t tend to capture a person when their eyes are closed. So naturally, the algorithm never inherits the learned behavior of how a person normally blinks, which produces a rather unnatural result. Since the AI detects how open an eye is in each frame of a video, it is able to catch when a person does or doesn’t blink.
Creating DeepFake videos isn’t all that easy to begin with. As the authors note in their study published in Cornell University Library, it takes a lot more time than simply photoshopping images together. For example, a 20-second video with 25 frames per second requires editing 500 images together. In fact, Buzzfeed spent a total of 56-hours creating that fake Obama video.
As we mentioned before, the technology continues to advance. Stanford researchers recently came out with a new algorithm that includes blinking, and they plan to present it at the SIGGRAPH conference this August.
The SUNY scientists add to New Scientist that “media forensics is a cat-and-mouse game.” Fake news has already incited violence, but the video element adds a much deeper issue. People tend to be more easily influenced and misled by videos than images.
“Detecting such fake videos becomes a pressing need for the research community of digital media forensics,” wrote the authors, who acknowledge that sophisticated forgers can add blinking in during post-production. The team soon hope to be able to pick up on other cues, such as breathing or a person’s pulse, that could indicate a fraud.
