What is a deepfake, and how can you recognize it?
Deepfakes are videos that have been doctored using artificial intelligence, and they’re a growing concern for governments, law enforcement agencies, and even individuals. The first step to protect yourself from such threats is education – learn how to recognize these doctored images and what potential harm they could bring to everyone. Let’s dive in.
What is a deepfake?
You can call this an evolution of Photoshop, which can be used to doctor images of real-life people in situations they typically wouldn’t be in. Deepfakes mislead audiences into thinking that the portrayed individual performed actions or said statements they never actually did, often for malicious purposes.
Deepfake definition
Deepfake refers to a form of synthetic media that employs artificial intelligence (AI) and machine learning techniques to produce convincingly realistic, altered audio and visual material. By merging or overlaying pre-existing images and videos, deepfakes mislead audiences into thinking that the portrayed individual performed actions or said statements they never actually did.
Deepfake examples
- Celebrities in fake videos: Falsely depicting a celebrity in a compromising situation, often used for defamation or as a tool for blackmail.
- Politicians in manipulated speeches: Creating videos of politicians saying things they never said, which can be used to spread disinformation or sway public opinion.
- Impersonation for identity theft: Using deepfakes to impersonate someone in a video or audio call to gain unauthorized access to sensitive information.
Deepfake vs. shallowfake
While deepfakes rely on AI to create highly realistic manipulated content, shallowfakes involve simpler editing techniques like speeding up, slowing down, or cutting and pasting video frames.
Pros and cons of deepfake technology
Pro:
- Deepfakes can be used for entertainment purposes, such as parody videos, or to create realistic visual effects in movies and video games.
Con:
- They can be used for nefarious purposes, including spreading disinformation, manipulating public opinion, and facilitating identity theft and cybercrimes.
Preventing deepfake deception
- Verify the source of the content before sharing.
- Look for inconsistencies in lighting, shadows, or audio quality.
- Use video forensics tools to detect manipulated content.
How do deepfakes work?
Deepfakes leverage deep learning and generative AI to create highly realistic content. The core of deepfake technology lies in using neural networks, a technique called a “generative adversarial network” or GAN. A GAN consists of two main components: a generator and a discriminator. They train themselves to recognize patterns using algorithms. The generator creates fake images or videos, while the discriminator examines them and blocks content it recognizes as fake.
The generator and discriminator engage in a continuous loop. The generator improves its output to trick the discriminator, and the discriminator enhances its ability to detect fakes. Over time, the generator produces highly realistic fake media that can be difficult to differentiate from authentic content.
The process typically starts with a large dataset of images or videos of the target individual. The deepfake algorithm analyses them to learn the subject’s facial expressions, body movements, and other nuances. The generator uses this training data to produce fake content.
What is an example of a deepfake?
Deepfakes showcase how AI can manipulate media in impressive ways. Imagine a video in which a celebrity’s face is seamlessly put onto another person’s body, making it appear as though the celebrity is performing actions or speaking words they never actually did. However, this is just one of the many applications of deepfake technology.
Textual deepfakes
Textual deepfakes involve using artificial intelligence to generate convincing fake text. This technology can create realistic articles and news stories or even impersonate someone’s writing style in emails and messages.
The AI analyses large datasets of text to learn and mimic the way a person writes or speaks. While this can be useful for automating content creation, it poses significant risks. People can use textual deepfakes to spread false information or deceive individuals into believing they are communicating with someone they trust, causing significant harm.
Video deepfakes
Deepfake videos use advanced AI techniques to create realistic fake videos by swapping faces or altering appearances. The process involves using neural networks to map and animate facial features, making the altered video appear genuine.
AI-generated videos pose serious ethical concerns. Criminals can use such videos to fabricate events, create fake news, or impersonate individuals, leading to misinformation, reputational damage, and potential legal issues.
Audio deepfakes
Deepfake audio technology uses AI to generate realistic speech that mimics a person’s voice. The AI analyses a target’s voice recordings and produces an audio track that sounds just like that person.
People can use this technology to create fraudulent recordings and impersonate others for malicious purposes. For example, scammers can use audio deepfakes to make fake phone calls and extract money.
Live deepfakes
Live deepfakes involve real-time manipulation of video and audio feeds to alter a person’s appearance and voice during live broadcasts or video calls. This technology can be used for virtual reality experiences, live performances, or interactive gaming, providing immersive and engaging experiences.
However, the potential for misuse is significant. Live deepfakes can enable impersonation during video conferences, spreading false information, or conducting fraudulent activities.
Deepfake nude bots
Deepfake nude bots use AI to create explicit images or videos of people by manipulating existing content without their consent. One notorious example is the Telegram deepfake bot. Researchers discovered that this bot forged nude photos of over 100,000 women. The bot’s AI algorithm replaces the clothed parts of an uploaded image with artificially generated nudity.
Cybercriminals shared the images online, mainly targeting women, with some being underage. Cybercriminals promoted the bot through various social media platforms, mostly Telegram and VKontakte.
Such bot cases raise deepfake threats to a new level because they involve private individuals and are dangerously accessible. If you have at least one photo uploaded somewhere online, you could easily become a potential victim of cyberbullying. Being alert online is crucial to preventing such threats. Limiting the content you share publicly can stop threat actors from using that content for malicious activities.
Is there a positive use case for deepfakes?
You can utilize deepfake technology for various positive purposes, such as creating realistic aged-up versions of missing people, realistic portrayals of historical events, or realism enhancement in gaming. However, producing and distributing deepfakes doesn’t require any form of agreement. Using another person’s image for content without their consent is ethically and often also legally wrong. The potential misuse of deepfakes can lead to misinformation, privacy violations, and reputational damage.
How are deepfake videos made?
Although deepfake AI technology is still evolving, it is already highly advanced and capable of producing realistic results quickly with minimal human intervention. Various free apps, such as Wombo and FaceApp, enable anyone to create deepfake media. While these programs may produce less convincing results, they represent the beginning of what is possible with this technology.
How to identify a deepfake
It’s not always easy to detect a deepfake. While some videos are clearly fake, with facial expressions giving off a surreal, uncanny valley effect, others are more sophisticated.
Several factors can help you determine whether you’re looking at a convincing deepfake or not. If the video contains a face, focus your attention there and look for these giveaways:
Smooth or blurry patches
The connecting points where the deepfake video overlay meets the face of the person underneath can sometimes appear oddly smooth and textureless. Even with better examples, any sudden head movements or changes in lighting can momentarily reveal blurry facial borders.
Inaccurate non-facial features
If the video represents a public figure, like a politician, you can find real images of that person to compare. Look at elements outside the main facial features that might not have been altered – hands, hair, body shape, and other details that don’t sync up between the video in question and older, more reliable visual sources.
Unnatural movement
While deepfake is a relatively new technology, in the last few years, it has improved a lot, and deepfake can now easily recreate a person’s movements. However, some details can still give a deepfake away.
If the subject’s body and head seem oddly stiff, it could be a sign that the video creators are trying to make it easier for the deep-learning AI to map an image onto the person’s face without having to track too much movement.
An unconvincing voice
Deepfake technology is rapidly evolving, but for now, training computers to create audio simulations seems to produce poorer results than synthesizing convincing deepfake images and video.
The creator of the deepfake has to choose between two options if they want their subject to speak – either use an AI-generated voice or an actor who can impersonate the source material. You can compare the voice to the audio of a famous person speaking, and you may notice some differences.