How to detect deepfakes in video calls

Last month a story broke that remote job interview scams involved impersonate someone else using deepfakes. And while the purpose of these actions has nothing to do with getting the job, it does present a new security issue.

The truth is that, to a greater or lesser extent, video calls seem to exist here, not only on a professional level but also on a personal level. Therefore, everyone is exposed to being deceived. The question is: can be easily detected deepfakes? The Metaphysical Company devised a curious technique.

A promising and simple tactic to detect deepfakes

A recent study claims that we “catch” less than half of the deepfakes that pass before our eyes, although of course, we do not always have the same probability of detecting them.

If a shared video is displayed on social networksunless users have specialized tools, they have no choice but to be attentive to try to determine if it is a hoax or not.

However, if a person is in a video call, the chances of detecting a deep deepfake increase. The other person may be asked to perform a series of gestures that falsify the functioning of the software that generates the deepfake in real time, and thus it can be detected trying to impersonate another person.

The team of Metaphysic asked actor Bob Doyle to participate in a series of video calls using DeepFaceLive, a fake real-time version of the popular DeepFaceLab software. As you can see from the screenshot, the system did a great job with most of the characters.

Doyle becomes Sylvester Stallone, Ryan Reynolds, Nicolas Cage, Keanu Reeves, Tom Cruise, Vin Diesel and others. However, in the middle of the show, he was asked to slowly turn his head to the side. Result? The quality of the deepfake is lost when the angle reaches 90 degrees.

As the team explains, most deepfake systems they don’t work well in sharp corner situations.

In fact, almost all of the benchmarks used by the facial recognition algorithm are assigned to the front of the face. So when someone shows his face on a fake profile, he starts to get confused.

One of the other techniques that Metaphysic offers to identify deepfakes is ask the person with whom you are making a video call to wave their hand in front of their face. This would create a “serious situation for the model that could show lag and poor overlay quality on the real face.”

As the technology advances and the algorithms are trained more, the deepfakes become more accurate and realistic. For now, however, someone can be asked to turn your head 90 degrees or what Wave at the camera to see if you’re lying.

This Artificial Intelligence detects deepfakes on TikTok

A team of researchers from Singapore has developed a system of artificial intelligence capable of determining in 98.53% if the analyzed video has been digitally manipulated.

This system has been developed by Wang Weimin, a scientist graduated from the National University of Singapore and who currently works at ByteDancethe parent company of TikTokwhich has entered a contest with hundreds of applicants organized by the ISGF, a national AI program launched by the National Research Foundation of Singapore.

The contest has lasted 5 months and Wang, who after winning received a prize of USD$300,000 to commercialize his idea, has assured that he prefers to implement his technology in his own company to combat manipulation in social networks.

Specifically, Wang wants to incorporate this technology able to detect manipulated images in BytePlus so that users can make use of it. This last mentioned platform is a subsidiary of ByteDance, specialized in the development and research around artificial intelligence.

“Good or bad, deepfake is an emerging technology that simply cannot be ignored”Wang commented after winning the contest. According to CNA, Singapore’s television, Wan decided to participate because she has a great interest in solving the great challenges facing the media.