“Donald Trump is a born idiot,” says the former president of the United States, Barack Obama, in a video from 2018. Obama himself has never uttered those words. The voice in the video belongs to the American comedian Jordan Peele, but the face and gestures are both made with artificial intelligence.
The video was prepared in order to demonstrate how easily it is possible to deceive the audience with so-called deep fakes, where live images and sounds are used to trick the audience with false information.
VIDEO: See Barak Obama Say Weird Things
In 2018, the American online news media Buzzfeed, in collaboration with the comedian, prepared a fake video in which Obama, among other things, called Trump a birth idiot.
Nowadays, deep fakes are far more common than ever and technology has advanced so much that it is impossible to spot fake videos with the naked eye. Technology companies now slow down every year to come up with deep fakes by applying their own algorithms.
Tech companies want to eradicate deep counterfeiting
Deep fakes are based on artificial intelligence that creates a digital replica of a person, for example Barack Obama, using countless photos, videos, as well as audio recordings, of the person in question.
Those behind the deepfakes can then get the victims in the videos to say and do almost anything, acting as a kind of digital puppeteer. Today, these tricks have become so developed that it is almost impossible to tell a fake from the truth.
It is worth noting, however, that people other than politicians end up being clones of the masters of deep forgery.
In 2019, criminals managed to embezzle 36 million Icelandic krónur from a British energy company by perfectly imitating the CEO’s voice in a deep fake recording.
A year later, the environmental organization Extinction Rebellion distributed a deeply faked video with the Belgian Prime Minister, in which she was supposed to have linked the climate problem to Covid-19.
Barack Obama has been popular among deep counterfeiters. Not least for the director and comedian Jordan Peele, whose video has gained worldwide attention. The technology behind the fake videos is constantly evolving and as a result it is becoming increasingly difficult to detect the fraud. The Barack Obama video can be used as a textbook example of how to detect deep fakes by focusing on three parts of the face.
Here’s how you can spot a deep fake
The Barak Obama video can be used as a textbook in detecting deep fakes by focusing on three different facial features.
1 – The eyes should blink
A so-called deep fake image is made up of countless images and videos that are put together in a computer so that a believable overall image is created. There may be slight deviations in eye movements that seem abnormal to the viewer’s mind, for example the eyes may suddenly look away from the camera or the person in the video never blinks. In the video showing Obama, these flaws have been corrected.
2 – Pixels should not be visible to opposite eyes
Due to the fact that deep fakes are often based on the fact that faces are replaced, there is a risk that irregularities in the periphery of the face can occur during the fraud. On Obama’s left eye, you can see the hair moving a little, almost in jerks, even if Obama himself is standing still.
3 – The mouth should become sensitive
If the mouth is in constant motion, flaws in the video are easily revealed, because the mouth movements cannot follow the words completely and as a result appear to have been synchronized. In many deep fake photos, it is similarly difficult to show teeth believably. In the video with Obama, there is nothing more like the fact that the teeth sometimes merge with the bottom lip.
If some of the world’s biggest tech companies have their way, deep fakes will soon be a thing of the past.
The big companies Google and Facebook have promised the value of hundreds of millions of Icelandic ISK to anyone who can root out and delete fake videos before they do damage.
It can also be mentioned that the employees of the Dutch company Deeptrace have succeeded in eliminating fake videos with an extremely simple concept: They use the technology behind the deep fakes against the fakers themselves.
Algorithm in the role of police
Deep forgeries are based on artificial intelligence based on two algorithms (algorithms) connected in a network, what in English is called a generative adversarial network (GAN) and they constantly fight against each other.
The second is the so-called creative algorithm. It is fed with and dissects a large amount of data about the person it is intended to impersonate in order for the deep fake to be as successful as possible. The algorithm separates facial features, hair, shadows, etc. from various perspectives and finally prepares a proposal for deep forging.
Algorithm number two then supervises the work and plays the role of a policeman. The algorithms come up with the wrong positioning of image pixels, the movement of the face relative to the head itself, and the like. If the monitoring algorithm notices a defect, it sends a corresponding message to the creative algorithm, which immediately corrects the defect and comes up with a new, refined proposal.
In other words, the creative algorithm is tested by the control algorithm, and when the control algorithm stops detecting a forgery, the deep fake is ready.
The little things make all the difference
Deepfakes first became widely known in 2017 when a certain user on the website Reddit started downloading sex videos. Soon after, users of the website were able to find videos in which well-known singers such as Taylor Swift and Katy Perry had been fraudulently inserted into the pornographic films.
Since then, thousands of deeply fake videos have been added to the Internet, from pornographic videos to videos featuring politicians. Now researchers and companies have joined forces to try to catch fake videos before they go viral.
One of these companies is the Dutch company Deeptrace. This company has managed to develop an algorithm designed to monitor video content.
The algorithm was fed with many thousands of hours of real and fake videos, and the algorithm was then designed to detect the differences.
If Henry Ajder, head of threat intelligence at Deeptrace, is anything to go by, they need to drill down to the smallest detail in order to spot fakes.
“The human eye has a hard time detecting many deep forgeries, and the defects are often hidden in the tiles themselves. Our algorithm is, among other things, trained to find tiles that are not in the right place, for example small tile pellets that are either too light or dark compared to the environment,” says Henry Ajder.
Deepfaking is done by combining several photos and videos, and when it’s done, there’s a risk that a few extra pixels will get thrown in. Therefore, the algorithm goes over every single pixel in all the images, often many thousands of pixels in each video, to make sure there are no inconsistencies, and they often come across something that doesn’t belong in the footage.
In December 2018, the company Deeptrace quickly found 14,000 videos online that they identified as deep fakes. Their employees have to work hard to constantly improve their knowledge, because the deepfake videos are constantly evolving:
“The masterminds are getting better and better at creating deep fakes, so we need to constantly refine the algorithms so that they can identify the fakes,” says Henry Ajder.
Gesture may be the weak spot
Those at the company Deeptrace focus on spotting flaws in the quality of the deepfakes: poor resolution, misplaced pixels, etc. but scientist Shruti Agarwal at the University of California has chosen to take a different route.
She noticed that Barack Obama moved his head slightly up, either to the right or to the left, every time he said “hey, everybody.” This led her to decide to investigate whether deep fakes could mimic how individuals speak and what facial movements are based on what is being said.
In 2019, Facebook founder Mark Zuckerberg was himself a victim of digital fakes, when a deep fake featuring him appeared on Facebook. In the video clip, Zuckerberg sits and admits that he controls the future and the lives of millions.
As a result, she used an analysis program to record the facial expressions of five different politicians, and this work has helped her uncover the weak spot in deep fakes. Deep fakes simply superimpose one person’s facial features on top of another person’s face, without considering facial expressions.
Shruti Agarwal’s algorithm was able to identify 94 percent of the deep fake videos she had the software look at . The solution lies in using quite a few pictures of the person’s gestures, and this means that the algorithm gives the best result when it comes to famous people because of the number of photographs of them.
The search has begun
The technology behind the deep fakes has developed so much that sophisticated algorithms have to be applied to spot the fakes. As technology advances, it becomes more and more difficult to spot deep fakes before it’s too late.
As a result, social media giants Google, Facebook and Twitter have launched initiatives aimed at finding and removing fake videos before they can spread the wrong information online.
Facebook has spent millions of dollars paying universities that work with image and video analytics and that conduct research on how to eliminate deep fake videos.
Google employees have recorded interviews with hundreds of actors and then used the recordings to prepare a total of 3,000 deep fakes. They then put the material on the Internet, where scientists and computer scientists used it to train the algorithms to come up with deep fakes.
Should it be possible to distinguish the truth from a fabrication in the future, it is time to eliminate all these fake videos. Or as the “fake” Barack Obama says in a deeply fake video from 2018:
“It seems simple at first glance, but our future behavior in this information age will determine whether we survive or end up in a bad place.”
