What is Deepfake? How to Detect Deepfake?

Abstract: The progress of deepfake technology seems to affect our perception of reality deeply. It is possible to create events that did not occur in reality by using this technology. In this context, some politicians saw themselves voicing words they had never said, and some celebrities found themselves appearing in various pornographic content even though they were not involved in it.

What is Deepfake?

Deepfake technology can embed any person in a video or photo where they’ve never actually been in a way that can’t be considered fake. The late Paul Walker’s reappearance in the Fast and Furious 7 movie was one of the examples of this technology in the past years. In the past, this kind of work used to take a whole year of big studios with many experts. But now, with self-running computer graphics programs and machine learning systems, deepfake technology can produce images and videos faster.

With the word “deepfake” creating a lot of confusion, computer vision and image processing researchers came together because of their hatred for the word. Because this word covers a wide area, from cutting-edge videos created by artificial intelligence to any photo that can be considered a fraud.

In fact, contrary to popular belief, some content is not called deepfake. An example of this is the famous “cricket” video featuring Michael Bloomberg, a past Democratic presidential nominee in the United States. This was actually created using simple video editing techniques, not deepfake.

How is Deepfake Content Created?

The main raw material of deepfake content is machine learning. Thanks to this, deepfake content can be created at less cost and much faster. To create a deepfake video of someone, a video of that person actually taking place must first be taken. Then the data from this video along with neural networks must be trained for hours. Only in this way, the images that the person should take under different angles and various types of light can be realistically imitated. Now, thanks to the processing of data trained by computer imagery methods, it becomes possible to copy one person onto another.

Although artificial intelligence shortens this long process, it still takes time to get a realistic result. There are also some manual operations that need to be done to portray a real person in a completely unrealistic environment. The spurious and obvious details that the trained data can generate have to be trimmed by hand, but the process is not that simple.

According to most people, contentious generator networks (GANs), one of the deep learning algorithms, will be one of the main methods used in the field of deepfakes. Fake faces created with GAN are indistinguishable from real human faces. The first inspection of deepfake applications was made using GAN algorithms. In this way, even anyone in the future can easily produce advanced deepfake content.

SEE ALSO: What is Deep Sleep (Hypersleep)?

Stating that some false assumptions have been made on this method, Siwei Lyu from SUNY Buffalo University says: “GAN algorithms do not play a significant role at all in most deepfake content produced today.”

GAN algorithms are difficult to work with because they need large amounts of training data. Unlike other methods, the result is much more time-consuming. But more importantly, while GAN models are good at rendering images, they are not as efficient when it comes to videos. Because while skipping the image frame by frame, they have difficulty in aligning it with the image in the skipped frame.

In addition, well-known voice deepfakes do not even use the GAN algorithm. An example of this is when Canadian artificial intelligence company Dessa (now owned by Square company) managed to use the voice of talk show host Joe Rogan to say some phrases that he had never actually said. GAN algorithms were not even used in this content. In fact, today’s deepfakes are mostly created using both AI and non-AI algorithms.

Are Deepfakes Just Videos?

Deepfake videos of famous names such as Mark Zuckerberg, Barack Obama, Morgan Freeman and Tom Cruise were on the agenda when they were published. Of course, deepfake technology is not limited to videos. Deepfake audio is also a rapidly growing field with many applications. Deepfakes can now be created using deep learning algorithms with just a few hours of voice from the person whose voice was cloned.

Another example is the sounds used in video games. Instead of relying on a limited set of scripts recorded before the game, programmers can now let game characters say anything in real time with this technology.

What Threat does Deepfake Pose?

Privacy and security advocates fear the technology will be used for fraud and blackmail; They say deepfake content will make it even easier. Experts state that deepfake videos emerge at a time when we are facing a media literacy crisis, meaning that the often harmful, misleading and false information spread by deepfakes means that many people will accept it more easily.

We should not confuse deepfakes with other video editing and photo editing software. The key to deepfake is the use of artificial intelligence and machine learning to create ultra-realistic videos. Therefore, it is very important to be aware of the deepfake threat and be able to detect them.

How to Detect Deepfake?

As deepfakes become more common, society will need to adapt to detecting videos and online users are now accustomed to detecting other types of fake news.

There are several indicators that give away deepfakes:

-Existing deepfakes have difficulty recreating faces realistically. Especially blinking is one of them. A video where the person never blinks, blinks too often or unnaturally, betrays a deepfake.

-You should look for skin or hair problems or faces that look more blurred than their surroundings. The focus may appear unnaturally soft.

-Lighting doesn’t look natural? Often times, deepfake algorithms preserve the lighting of clips used as models for fake video; a poor match occurs with the lighting in this target video.

While deepfakes will become more realistic over time as techniques improve, we are not entirely vulnerable when it comes to tackling them. Some companies are developing methods to detect deepfakes simultaneously.

Source: https://mitsloan.mit.edu/ideas-made-to-matter/deepfakes-explained