Deep fakes attract a lot of attention in those days. Depending on the interpretation and technological understanding, we are on the verge of losing a lot of control over our face and our voice, or we are facing the next logical step after the possibilities of changing static images using Photoshop and other image processing software.
What are Deep Fakes?
The term consists of “deep learning” and “fake” — the learning of artificial intelligence with the intention of creating a targeted fake. In most cases, this is based on video manipulations in which software analyses the source material and extracts part of it, then inserts and adapts it in another video. So-called “face swaps” are the most common form of Deep Fakes — i.e. the swapping of faces.
But development continues: Adobe recently introduced the Cloak project, in which intelligent technology helps to remove specific content from a video.
How to create Deep Fakes?
Whereas in the past it was necessary to use expensive software to change moving image material and this was therefore mainly reserved for film production in Hollywood, today every user can do this himself with the corresponding apps on his smartphone. The result is currently still an approximation, but even with a more elaborate graphics process and the corresponding (freely available) software, quite professional-looking deep fakes can already be developed. To generate “good” deep fakes, you still need about 300–2000 images as source material, which the artificial intelligence has to analyse to learn from.
How common are Deep Fakes?
The phenomenon first appeared to the general public about a year ago. At that time, Motherboard reported for the first time on deliberate fakes in which the faces of celebrities were incorporated into porn videos. The technology received worldwide attention after researchers at the University of Washington put fictitious words into US President Barack Obama’s mouth. At first glance, this involves many dangers and countless abusive use cases. However, so far no case is known in which a deep fake video was deliberately used to manipulate political processes or even to falsify elections.
How to discover Deep Fakes?
Prof. Ira Kemelmacher-Shlizerman of the University of Washington argues that if it is known how such deep fakes can develop, one can also develop a technology that can detect such deep fakes. This counter-development is still in its infancy: two researchers from New York State University in Albany have developed a learning system in an experiment with around 25,000 images that should detect deep fakes in the future.
In addition to research, industry and the US military are also working on developing software-based solutions to combat deep fakes: the GIF platform Gfycat recently announced that they had developed a tool based on artificial intelligence that will detect and block deep fakes in the future. Facebook also states that they want to use a tool they developed themselves to track down deep fake videos in a first step so that they can then be checked manually by fact checkers.
However, all of these technology-driven aids against deep fakes have one thing in common: they need a lot of source material in order to systematically build up the necessary intelligence. This is still time-consuming and takes a lot of time at the moment. Therefore, in the medium term, journalistic verification of these contents is still the only realistic way to fight deep fakes.
How to verify Deep Fakes?
As always, verification is a multi-step process that is similar to a puzzle game. Here are 8 steps to verify Deep Fakes:
1. flickering faces
Let’s start with the obvious: flickering faces. In many deep fakes, the faces still look strange. The transitions between the face and the neck or the hair do not always fit together. If everything else looks normal, but the face looks strange, it is probably a fake.
2. face and body
Also obvious, but perhaps neglected at first glance: does the body fit the face or does the posture match the facial expression? Most deep fakes are primarily facial substitutions — changes to the body can only be implemented with great effort. If the person is noticeably heavier, lighter, bigger or smaller and suddenly has other conspicuous features (e.g. tattoos, pronounced muscles, different skin colour), this also indicates a fake.
3. length of the clip
Short clips. Although the technology is already very easy to use, the learning processes for making deep fakes are still laborious. Therefore, most deep fake clips that are shared are only a few seconds long. So if a very short clip of implausible content is to be verified and there is no obvious reason why the recording is so short, then this is often an indication that it is a fake.
4. source of the recording
In this context, searching for the source of the recording — the person or account who first shared the video — also helps. Often this helps to find out the context of the publication and to check if the source material was more detailed after all.
5. sound for recording
Not only the picture exposes deep fake videos, but also the sound. Deep fake software is often limited to changing the image, but not to adjust the sound. So if the sound is not present or does not match the image — for example with poorly implemented lip synchronization — this again indicates a fake.
6. recognize details at half speed
Note the details: When verifying video content, it is also helpful to let the video play at half speed. For example, small discrepancies in the background of a person or sudden changes in the picture are noticed more quickly.
7. fuzzy interior of the mouth
Software for creating deep fakes has been able to transfer faces pretty well so far, but the devil is in the details. For example, a certain blur in the inside of the mouth is another indication that it could be a fake image. Artificial intelligence is currently still struggling to correctly represent the teeth, tongue and oral cavity when speaking.
8. note blinking
Another hint may be the blinking of the speaking person: Healthy adults blink every 2–8 seconds. This blinking can last between a tenth and a quarter of a second. You could assume this to expose fake videos because most software can’t yet make the blink normal.
Which tips for the verification of Deep Fakes do you know? Write it to me in the comments or contact me via Twitter (@konradweber) — I am looking forward to your answer.