What is deepfake and how to detect it?

Deepfake uses deep learning artificial intelligence to replace one person's likeness with another in video and other digital media. One of the most obvious manifestations of what is now called "synthetic media" is images, sounds and videos that appear to have been created using traditional methods, but are actually created with deep learning and artificial intelligence using complex algorithms. Most of the time, content created with this technology is indistinguishable from reality.

How Deepfake Technology Works

“Deepfakes” refers to the technology underlying “deep learning,” a type of artificial intelligence. With the help of deep learning algorithms that learn how to solve problems based on large amounts of data, fake media can be made to look realistic.

Deepfakes can be created in various ways. One of the most common includes deep neural networks and autoencoders that use the face swapping method. The first thing you need is a target video for the deepfake, as well as a collection of clips of the person you want to target. Videos may be completely unrelated; the target may be a clip from a Hollywood movie, and the videos of the chosen subject may be random clips from YouTube.

An autoencoder is an artificial intelligence program created to examine video clips to determine how a person looks from different angles and in various weather conditions. It is then used to find similarities and match that person to the person in the target video.

How to Use Deepfake?

There are some interesting applications (as in movies and games) of automating the swapping of faces to produce synthetic videos that look reliable and realistic. That's because deepfake technology was first applied to create synthetic pornography. In fact, according to Deeptrace, 96% of deepfake videos found online in 2019 had pornography as the main content.

Since then, the technology has been developed for use in leading national figures.

Is Deepfake Just Video?

Deepfake technology is not limited to videos. Audio is a rapidly growing field with a wide variety of applications.

With deep learning algorithms, realistic voice spoofs can be made from only a few hours (or in some cases, minutes) of the voice of the person whose voice is cloned. After a voice model is made, that person can be made to say everything.

How to Detect Deepfake?

As deepfake becomes more common and online users now gain experience in detecting other types of fake news, society is likely to adapt to detecting deep fake videos. But in cybersecurity, detecting and preventing deep fraudulent technology often requires more innovation.

There are several indicators that allow deepfake to be detected:

·       Existing deepfakes have difficulty recreating faces realistically, resulting in videos where the person does not blink at all or blinks too often or unnaturally. But after University of Albany researchers published a study that detected the blinking abnormality, new videos were released that didn't have this problem.

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

·       Does the lighting feel natural to you? Deepfake algorithms typically preserve the lighting of clips that are used as models for fake video and that do not match the lighting in the target video.

·       In some cases, the audio may not match the person, especially if the video is fake but the original audio has not been carefully modified.

Combating Deepfake Technology

As techniques improve, deepfake scams will become more realistic, but you're not completely vulnerable when it comes to fighting them. Some startups have developed and continue to develop methods to detect these scams.

For example, Sensority has developed an antivirus-like platform for deepfake videos that alerts users via email when watching something bearing the obvious fingerprints of artificial intelligence-generated synthetic media.

Operation Minerva identifies deepfakes in a simpler way. Operation Minerva's algorithm compares potential deep frauds with known "digitally fingerprinted" videos. For example, Operation Minerva may identify that a video it has already cataloged is a modified version of it.

Deepfake Video and Audio Detection

AWS, Facebook, Microsoft, the AI Media Integrity Steering Committee and academics came together to create the Deepfake Detection Challenge (DFDC). The goal of the challenge was to encourage researchers around the world to create innovative new technologies that can help detect deepfakes and manipulated media.

Because as this technology advances, it will become more and more difficult to determine what is real and what is not. That's why it's so important that we don't trust and verify what we see online before posting any video on social media.

Experts predict that as technology improves, deep frauds will become more sophisticated and pose a greater threat to people, such as election interference, political tensions and different criminal activities.