Get Avast One to help block scams and protect your phone
- Security
- Privacy
- Performance
Deepfake videos blur the boundaries between reality and fake news, making it increasingly difficult to believe what you’re seeing. But if you know what to look for, there are telltale signs to betray these unnervingly real and often bizarre videos. Learn how to spot a deepfake — then stay safe online with a comprehensive security and privacy tool like Avast One.
A deepfake is a deceptively realistic but fake piece of media created by altering existing video or audio material. Deepfake videos use artificial intelligence (AI) tools to convincingly replace a person’s face or voice with somebody else’s, allowing the synthesized media to be used playfully or to spread misinformation maliciously.
This Article Contains:
Fake digital media is nothing new, but deepfakes are a thoroughly modern phenomenon. Unlike fake pictures created by real humans using tools like Photoshop, deepfakes are fabricated by AI machines. Use of this kind of sophisticated video manipulation is increasingly common in pharming schemes, making it harder to protect yourself against identity theft.
The word deepfake is itself a synthetic creation — a portmanteau of “deep learning” and “fake.” The deep part of the deepfake meaning refers to deep learning, a method of training computers to think naturally like a human brain. The fake part of the definition emphasizes the deceptive nature of deepfake media.
Deep learning is an AI process that involves repeating a task over and over again, sometimes without any human supervision, in order to discover the best way to deliver a desired outcome. In the case of deepfakes, the AI is fed hundreds of reference photos and videos to teach it how to generate a version of a person’s face that can be animated.
Deepfake videos are often used for harmless purposes like in memes, social media filters, or face-swap apps. But deepfakes can also be used maliciously to spread misinformation, create fake news, or launch revenge videos. Altering someone’s appearance to make them look like someone else is often done playfully — but it can also be used for cyberbullying or lead to identity theft.
Indeed, it was the seedier side of deepfake technology that first burst into the public domain. In 2017, a Reddit user originally used the technology to digitally create non-consensual porn using female celebrities’ faces. While the subreddit was banned in 2018, this type of Reddit deepfake, and other forms of celebrity deepfakes, have spread across the internet.
And that’s just the beginning when it comes to notorious deepfake examples. Deepfake technology can be used for many types of deception, from political manipulation and fake news, to revenge porn and blackmail. Anyone with access to deepfake technology can make anyone else look like they are saying or doing pretty much anything.
The danger of creating and weaponizing deepfake videos is already upon us. And there are growing concerns that deepfakes will increasingly be used maliciously in the future, calling into question the security of biometric data, including in the use of facial recognition tools.
Deepfakes are created by feeding lots of images of a subject to deep learning computer networks, called variational auto-encoders (VAE). The aim is to train the VAEs to capture an array of lighting exposure, positions, and emotional expressions. That helps the AI determine which visual elements like expressions or shadows are unique and which are replaceable.
Here’s a closer look a the process of creating a deepfake video:
First, the AI needs two groups of input images: the original source (person A) and the target of the deepfake (person B). Computers can be trained on an array of random faces or on many images of a specific person.
Then, the AI creates the output images. The AI determines which subtle elements of expression are unique and essential to the deepfake. To be convincing, the AI must retain subtleties of personality and facial movements that occur naturally and are unique to the target individual.
To execute the face swap, the input and output images are combined — the original source’s output data is fed into the VAE and merged with the data of the deepfake’s target face. The encoder reconstructs the movements and emotional expressions of person A with person B’s face. This must be done frame by frame for a deepfake video to be truly convincing.
Yes, you can create a deepfake of someone’s voice using AI-enabled deepfake technology that uses machine learning and text inputs to create a synthetic deepfake voice that closely mimics that of a real person. Like other AI deepfakes, voice deepfakes can be made with popular apps and other online tools.
The best examples of deepfakes are usually those specifically created to demonstrate the power of deepfake technology. Other, more playful deepfakes are a bit less sophisticated but still interesting.
This leaves some people wondering what's so bad about deepfakes? When used as a tool for fraud, manipulation, revenge, or deception — particularly to misinform or persuade political views — deepfakes can cause a lot of damage on both a personal and societal level.
Check out the following deepfake example of former US President Barack Obama where he uncharacteristically drops the F-bomb and curses Donald Trump. One VAE was trained on input from actor Jordan Peel, the other on hundreds of pictures and videos of Obama.
This deepfake of Jon Snow apologizing for the Starbucks-cup cameo on the final season of Game of Thrones also shows the humorous side of synthetic videos.
This Joe Biden deepfake voice — made using a popular and easily-accessible voice deepfake generator — shows just how easy it is to create convincing AI deepfakes with today’s online deepfake technology.
Nixon’s failed moon landing deepfake speech is another convincing example of how easy it is to make it look as if political leaders have said or done something they didn’t. One of the reasons it seems so authentic is because the text is based on an actual speech prepared for Nixon if the Apollo 11 mission ended in failure.
Unsurprisingly, celebrities are often the subject of deepfakes. The fact that they’re well-known makes them ideal targets for memes, and the wealth of easily accessible media helps AI technology create quite convincing deepfakes of them.
While the term celebrity deepfakes may forever be associated with the pornographic Reddit deepfakes scandal, celebrity deepfakes are often lighthearted and transparent in their uses. Check out the following compilation of Tom Cruise deepfake videos, or this obvious mockup of a deepfake Trump reindeer story.
One of the best examples of a celebrity deepfake is this deepfake poem by Impressionist Jim Meskimen. With the help of deep learning AI, he seamlessly morphs from John Malkovich to Robert De Niro, and Arnold Schwarzenegger to George Bush.
Although superficially convincing, deepfakes can be distinguished from real videos quite easily if you know what to look out for. Despite the ability to produce very realistic AI generated face-swaps, convincing animation is another story, and there are almost always subtle signs that render something amiss.
But, as the technology matures and improves, detecting deepfakes is likely to become more difficult. Keeping the following deepfake detection tips in mind may help you avoid scams like phishing attacks or falling for fake news.
The biggest deepfake giveaways are audio flaws, awkward shadows and skin tones, and soft or blurred areas during movement. Pay extra attention to the mouth, forehead, and neck, as these are areas of the body that can often betray the hoax.
Listen to this deepfake of Ukrainian president Volodymyr Zelensky supposedly calling on his troops to surrender. When compared to an authentic clip of Zelensky, the altered voice sounds subdued and mumbled. You can also spot a deepfake by noticing audio that’s out of sync compared to the physical movement of the mouth.
Shadows and skin tones are incredibly difficult to render authentically, making them some of the more common ways to spot a deepfake video. In this Home Stalone clip, Kevin’s face often looks as if it literally has a Stalone mask on due to poor skin tone integration. You can easily trace the line from Kevin’s forehead, down his cheek, and across his jawline.
Soft or blurred areas, particularly around the mouth, are another dead giveaway that you’re dealing with a deepfake. The Tom Cruise deepfake above hits a lot of convincing angles, but if you watch closely, you can see soft, blurry areas around the chin, which become most noticeable during facial transitions and movement.
Deepfakes per se are legal almost everywhere, but the legality of specific deepfakes depends on context and intent, and varies from country to country.
In the US, specific laws are gradually being put in place to regulate deepfakes. Most US states have laws against revenge porn, but only a handful, including California and Texas, specify deepfakes as an illegal medium of such. California has also banned the use of deepfakes of government officials and candidates during election campaigns.
The UK does not currently have any specific deepfake laws. That means people maliciously targeted by deepfakes are left to use existing laws, such as those against defamation, to bring cases to court.
China recently enacted a deepfake law that puts liability on the platform itself for spreading deepfakes, and forbids platforms from recommending synthetic content. The law refers broadly to false information, including deepfakes, fake news, and other media deemed to be deceptive.
While strong cybersecurity protections exist against malware, hackers, and other traditional threats, deepfakes pose a new type of challenge. Spotting a fake photo or video online is not always easy, and deepfake detection will become increasingly difficult.
But security and tech companies are already working on innovative ways to combat malicious deepfakes. Schemes and tools like DARPA and Facebook’s Deepfake Detection Challenge are being used to arm society with a defense against false information. It's also important to know the differences between privacy and security so that you can defend yourself in both spheres.
Because deepfakes are created using large volumes of photos and videos targets and subjects, limiting the amount you share online — particularly on social media — can help you avoid becoming a target of a deepfake. Make sure you know how to change your Facebook privacy settings and how to find and remove personal information from the internet.
Deepfake audio and video could easily be used for phishing attacks, so stay aware of this new threat as it continues to develop. As with other forms of phishing, applying extra scrutiny to communications can help you catch fraudulent messages. If you’re unsure whether a video is real or a deepfake, do some research — Google it or see if you can find the source of the video on social media or elsewhere.
Once sensitive data falls into the wrong hands, there’s no knowing what a cybercrook might try to do with it. Your personal photos, videos, and information belong to you — so keep it that way! Protect your privacy online with all-encompassing privacy and security software like Avast One.
Avast One covers all your angles with an award-winning antivirus, an ironclad firewall, VPN and anti-tracking tools for anonymous browsing, and 24/7 data-breach monitoring to alert you if any sensitive account info is leaked online. Safeguard every aspect of your digital life today.