Share This Article
You shouldn’t believe everything you see on the internet, especially now more than ever. Advancements in artificial intelligence technology allow anybody to create fake audio and video content, known as a deepfake. And the scary part is that deepfakes look convincing.
If you are curious about what a deepfake means or fancy a high-level understanding of the technology behind them, this article covers everything!
What is a Deepfake?
A deepfake refers to any picture, audio, or video clip digitally altered to generate realistic-looking evidence of fake events. In the most common use case, people create deep fakes to replace one person’s face with another.
Deepfakes were popularly used to face-swap porn stars with female celebrities and add them to porn scenes. In fact, according to a Dutch startup firm, Deeptrance, 96% of deepfake content generated in 2019 was pornographic.
People continue to develop more user-intuitive tools that allow the less tech-savvy to create deep fakes needing fewer pictures and videos of the target person. Many fear bad actors will abuse this technology to make revenge porn and carry out other illicit acts.
Deepfake videos first became popular in 2018 when Jordan Peele, a famous comedian & actor, uploaded a disturbing video of Barack Obama. The video begins with the former American president speaking. Halfway in, the clip cuts to a split screen showing Jordan Peele matching Barack Obama’s facial expressions and lip movements. Peele made this video to show people how easy it is to spread false information online.
As demonstrated in that video, someone can use deepfakes to put words into people’s mouths by training one person’s facial expression on another and impersonating their voice.
Similarly, the images of other public figures like Tom Cruise, Donald Trump, and Facebook CEO Mark Zuckerberg have been subjected to viral manipulated media.
In the next segment, we’ll discuss how deepfakes work under the hood.
How Does Deepfake Technology Work?
The term ‘deepfake’ is coined from the artificial intelligence algorithm that makes them possible: deep learning. Deep Learning is a kind of machine learning that programs a computer to study a large set of data and solve complex problems using deep neural networks.
Put simply, deep learning is a sophisticated program that teaches computers to learn new things like humans through several examples.
Although deep fake technology uses numerous machine learning methods to create doctored videos, the most common technique is ‘face swapping.’
Let’s see how that happens. Heads up, we are about to get a little technical.
Steps to Create a Deepfake
To create fake videos, you first have to feed an AI algorithm called an autoencoder with some source material. You provide it with many photos of the person to be faked—preferably with many different perspectives of their face from differing angles. So things don’t get confusing, let’s call that the ‘target face.’
In the same way, you run several images of the person you wish to replace the initial target’s face through a separate encoder. Let’s call that the ‘replacement face.’
Typically, the quality of deepfake videos depends on the number of photo streams provided. The more images fed to the algorithm, the higher the fidelity of the results.
The encoder compares the two sets of images, reduces each to their similarities, and outputs a compressed image. The output of the encoder is called a ‘latent image.’
Both encoders have their respective decoders for recovering the original footage. The decoders have a reversed function of the encoder—they revert the ‘latent image’ to its original state.
Finally, you get your deepfake video by feeding the latent image to the technically ‘incorrect’ decoder. So instead of feeding the replacement’s latent image to its own decoder, you run it through the target’s decoder.
Although most people think deepfake video techniques employ the use of Generative Adversarial Networks (GAN), that’s an earlier method that has now become obsolete.
Deepfakes have become so convincing that it’s hard to tell the difference between the original and deepfake video with an untrained eye. Sophisticated machines with high computing power create painfully-accurate fake content that even passes through deepfake detectors unnoticed.
What Are the Different Types of Deepfakes?
Although manipulated video clips are the most common type of deepfake creations, other media formats like photos and audio recordings get digitally altered.
Deepfake Video
These are fabricated videos generated with artificial intelligence with the intent to mimic reality.
While the benign use of deepfake technology to make internet memes and face filters has been received as playful, its unethical application to produce and distribute harmful content has sparked controversy and fear.
With the accuracy of fake videos nowadays, one can see them being used to perpetuate all kinds of criminal and nefarious acts like identity theft, cyberbullying, and advanced phishing schemes via a convincing video call.
Creating deepfakes to put words in the mouth of political figures is an even more significant concern. Malicious users can puppeteer fake videos to influence voters, manipulate stock prices, and ultimately spread fake news.
In October 2017, deepfake videos appeared on the internet for the first time. A Reddit user uploaded altered pornographic videos that had the face of porn stars switched for female celebrities.
Although a deepfake creation that has undergone multiple rounds of re-rendering can be highly deceiving, some telltales may be a clear giveaway.
Typically, skin tones and shadows are difficult to render with AI software. Other telltale signs are blurred areas that pop up whenever the subject moves.
Deepfake Photos
It’s not only videos. Deep fake technology can generate portraits of people that don’t exist from scratch. There have been multiple sightings of online user profiles created using deepfake photos of people that don’t exist in the real world.
There’s reason to believe that one Maisy Kingsley, who posed as a Bloomberg journalist on LinkedIn and Twitter, was a fictional character.
Although deepfake images take significantly less rendering time and processing power than it takes to create fake videos, they can be harder to spot.
AI-powered tools designed with a generative adversarial network are commonly used to create doctored images. The AI algorithm has a generator that generates human faces with random facial features when fed noise.
A world-renowned deepfake researcher, Siwei Lyu, and his fellow scientists at the University of Buffalo invented a tool for detecting deepfakes. According to Siwei Lyu, the tool has an accuracy of 94% for detecting deepfakes with images in portrait format.
The tool spots fakes by comparing the reflection of light on the cornea of a human eye in photographs. Images found to have non-symmetric light reflections fail the test.
Deepfake Audio
Deepfake audio is an auditory recording processed with AI-powered software to generate compelling impressions of people’s voices.
Audio deepfakes are innovative tools with vast commercial applications. Authors, for example, use synthetic recordings to generate audiobooks of existing books. Podcasters can also use this technology to reproduce a certain voice quality lost to throat diseases and other medical conditions.
Other applications in the film industry improve the audio quality of dubbed movies. In addition, auditory deepfakes have the potential for more advanced utilities like personalized virtual assistants.
Although synthetic auditory media positively impacts lives, we can’t ignore their side effects on society. Easy access to deepfakes by unskilled users via smartphones and desktop devices has made them tools for spreading false information.
What Are Deepfakes Used For?
Deepfakes could be used for resourceful, playful, or ugly applications. It all depends on the intent of the user. While some may manipulate digital content for laughs, there is hateful deepfake content created to muddy peoples’ reputations and achieve selfish motives.
Parody & Humor Fake Videos
Deepfakes often circulate on the web for sheer humor. While people find it ridiculous to reduce such sophisticated technology to trivial use, you can’t deny their entertainment potential.
Social media platforms like Snapchat and Instagram take advantage of the potential for deep fakes to augment reality by adding playful face-swapping filters to images and videos.
There are face swap applications available on smartphones and PCs that allow users to replace the likeness of people with someone else’s. Some of these face swap apps are trained with thousands of images and video clips of certain celebrities, making them capable of producing realistic fakes of famous people. Just for kicks, users enjoy switching out their favorite star’s faces.
Illicit use of Deepfake Videos
Unfortunately, the dark side of deepfake digital content dominates the interwebs. A good ratio of fake media produced and consumed online constitutes non-consensual porn starring the faces of famous actresses.
Bad actors weaponize deepfakes created to perpetuate all sorts of nefarious actions. To name a few, people utilize this media for:
- Spreading hoaxes and fake news
- Cyberbullying and cybercrime like phishing, identity theft, and spoofing
- Political manipulation, e.g., Influencing voter decisions, misinformation campaigns, and manipulating stock prices.
- Other bizarre uses like revenge porn and blackmail
It is important to raise awareness about the notorious use of synthetic media to protect unsuspecting people from falling victim. People must learn to stop taking online content at face value, spot potential fakes, and double-check facts to control the spread of false information.
How To Spot Deepfakes
For the time being, the easiest way to spot deepfakes is to look closely for any auditory flaws, misplaced shadows, patchy skin tones, and flickering around certain areas during movement. Areas to watch are the forehead, mouth, and neck.
Although there is software designed to detect deepfakes by finding digital flaws, you can train your eyes to find gaps mechanically. No matter how smoothly someone renders a fake, you can detect media manipulation if you know what to look for.
Finding Audio Flaws
Generally, deepfakes have poor sound quality. Compared to standard media content, the subject’s speech in deepfakes sound muffled, as if passed through a low-pass filter.
Sometimes, the words and movement of the speaker’s lips are not in sync. But that’s not a very good pointer since lagging, or leading audio is a common technical difficulty with ordinary video clips.
Misplaced Shadows and Inconsistency in Skin Tone
Sketchy shadows and skin tone are one big indication that there’s something amiss. Since those are the most complex parts to render in a deep fake, they often betray the video.
When the subject’s shadow doesn’t align with the direction of light and the position of their body, you can suspect that someone tampered with the video. Uneven tones in the skin are another giant red flag. Sometimes, it seems as if the person in a fake video is wearing a face mask. An area around the subject’s face appears to be a different color from the rest of the body.
Noise or Glitches
The noise around the chin of transposed faces is a surefire way to tell that you’re looking at a fake. They become obvious when the subject moves their head around or makes facial transitions.
If you ever find yourself doubting the authenticity of a video call, making the caller turn their heads sideways can expose a potential deep fake. Besides the noticeable glitches during movement, side profiles don’t render with the same level of accuracy as the front view.
Implications of Deep Fakes
It’s almost impossible to imagine a world where deep fakes won’t lead to more mischief. Harmless and resourceful use of this technology is undeniably beneficial for improving the quality of people’s lives. However, its malicious application is what we should be worried about.
The real question is how much mayhem can deep fakes cause. Does it have the potential to instigate World War III? Perhaps a deepfake of a powerful nation declaring an invasion of a neighboring country.
The short answer is no. Developed nations have military-grade technology that can detect fakes in one breath.
But we cannot rule out mischief-making on more minor scales. The public eye is constantly tuned to and extremely sensitive to the media. Attempts to distort messages, take things out of context, or outrightly invent fake events to serve malicious motives can quickly mislead people and stir mass reactions.
Less obvious repercussions revolve around the growing distrust for media content. People already think of TV and the internet as agents of mind control and mediums for spreading lies. Repeated exposure to deep fakes will eventually numb people into indifference as it becomes more difficult to tell what’s real. As a result, people can strategically exploit this weakness to stir doubt about actual events.
Summary
Given its ability to create synthetic media, deepfake technology can be a gnarly tool in the hands of malicious actors. There’s always the fear of people using a convincing deepfake to perpetuate political agendas like smear or disinformation campaigns and spread fake news.
However, deepfake created for good use can make the world a better place. This technology improves the performance of many sectors, especially the entertainment industry.
