A portmanteau of “deep learning” and “fake,” “Deepfake” is a catch-all term for any media created using Artificial Intelligence (AI), Machine Learning (ML), or a combination of both.
Leveraging the concept of “deep learning” – a Machine Learning concept—deepfakes enable the creation of images, videos, and audio with a high degree of realism, often closely mimicking voices, tonality, and mannerisms in an extremely convincing manner.
How are Deepfakes Created?
Deepfakes are inherently reliant on deep learning algorithms, typically those found in Generative Adversarial Networks (GANs). A GAN is comprised of two parallel neural networks: the generator and the discriminator.
The generator is responsible for creating the actual content, whether it’s an image, audio, or visual. After that, the discriminator acts as an auditor, conducting evaluations on the authenticity of the content. Over time, the generator can learn based on user inputs, and its ability to create increasingly realistic content is augmented to a level where even the discriminator cannot detect it as a deepfake.
Ultimately, once the content is impossible for the discriminator to flag as AI-generated, it is deemed ready to be used in various Social Engineering attacks, such as Phishing, scams, identity theft, etc., with a high probability of compromising its human targets.
History of Deepfakes
Deepfakes started to emerge in the early 2010s following breakthroughs in Deep Learning. They didn’t gather mainstream attention until 2017-2018, when notable celebrities were imitated in AI-generated videos and circulated wildly, prompting attention from celebrity media outlets as well as technology and mainstream news.
While the focus of these initial deepfakes was for malicious purposes (explicit, nonconsensual content) or humorous videos (public figures making outrageous or erroneous statements), the use case for deepfakes rapidly expanded. Following advances in deep learning technology, threat actors shifted their attention to the specific targeting of education and national security and began to augment their social engineering and phishing capabilities with deepfakes.
What are deepfakes used for?
Deepfakes are not inherently malicious. For example, entertainment firms and production houses leverage deepfakes to create special effects, de-age actors, or even portray deceased actors in cinematic scenes.
In the education sector, deepfakes can help with visual learning. They can generate historical reenactments and even create virtual tutors for students to clarify queries. From a marketing standpoint, deepfakes can often be leveraged to offer personalized campaigns. Tech firms are also increasingly using deepfake tech to drive advancements in AI and machine learning.
The reason deepfakes are notorious, however, is due to the scope for them to be used maliciously. Deepfakes are heavily used in identity theft, fraud, bypassing biometric security, the creation of fake news, and ongoing efforts to undermine public trust in mainstream media. Notably, deepfakes have been observed being used in defamation, election, and political tampering via the creation of fake content that can influence public opinion.
Different formats of deepfakes
Deepfakes have an extremely wide scope, being able to take the form of text, video, audio and other content to create highly sophisticated and convincing media to better compromise the intended target. Some common forms that deepfakes take are:
1. Videos:
Deepfakes often take the form of AI-generated but realistic videos, swapping faces, creating artificial characters, or editing existing footage. They are commonly used in movies for special effects, and threat actors can leverage these for malicious purposes by creating fake videos for blackmail, defamation, and even character assassination.
2. Audio:
Generating voices or altering existing voices using AI is a very common application of deepfake technology. Deepfakes can realistically create a convincing speech that mimics specific individuals. They can be used for benign purposes, such as virtual assistants or entertainment, but like most technology, they can also be used for impersonation, fraud, or misinformation.
3. Images:
Generative AI has greatly augmented the ability to create images, which are typically used artistically for visual effects, advertising, memes, etc. On the flip side, these are often used maliciously via the creation of fake images to conduct misinformation campaigns or to blackmail targets.
4. Text:
Deepfake texts are heavily used in chatbots as well as for content generation. Again, there is potential for abuse by threat actors using these for creating fake news content or misinformation via social media.
How to spot a Deepfake?
As deepfakes grow more sophisticated and lifelike, they can be tricky to spot, but there are several giveaways that can help determine if specific content is a deepfake or not:
1. Unnatural Facial Movements:
In videos, a common giveaway for deepfakes is unnatural or static expressions, such as the eyes and mouth not syncing or unsynchronized blinking. Look for these key giveaways on any video you suspect is a deepfake.
2. Lighting inconsistencies:
Deepfakes can often have lighting inconsistencies in the final output, particularly where shadows and reflections are involved. Elements such as light, shadow, or reflections may not naturally align with the rest of the frame.
3. Blurred or Flickering Edges:
When it comes to faces, keep an eye out for blurred edges, which indicate an overlay of an AI-generated face onto a background, and other indicators such as flickering.
4. Audio-Visual Mismatch:
Deepfakes can often have a desync between the audio and video, so one good way to identify a deepfake is to look for accents, lip movements, and tonality that does not match the speaker’s facial movements.
5. Background fluctuations:
In image or video deepfakes, the background could be distorted or not sync with the person’s movements, indicating a potential deepfake.
6. Unrealistic or “perfect” portrayals:
Deepfakes can often appear unrealistic or flawless, with regard to skin or overly perfect features that can give it away as artificially generated.
7. Deepfake Analysis Tools:
In the wake of high-profile deepfake attacks, specialized software and tools have been designed to detect deepfakes by analyzing any inconsistencies in the video or audio data, making the identification of a potential deepfake much easier.
Risks of Deepfakes
Deepfakes pose several risks, particularly given the speed at which AI and Deep Learning technology are evolving. The potential risks of falling prey to a deepfake attack can include identity theft, which for a compromised individual in a position of power can lead to massive consequences if they’re in a position of power in their organization or hold a position in the government.
More commonly, Deepfakes facilitate scams such as fraud, cybercrime, extortion, blackmail, and misinformation, which can have far-reaching consequences for national security, harmony, and the economy.
Deepfake FAQs
How is a deepfake different from Photoshop or Faceswap?
Photoshopped or face-swapped media is manual in its creation, requiring a high degree of skill and time to make a convincing image or video. Faceswap, on the other hand, uses relatively simple algorithms to digitally replace faces in media.
Deepfakes are entirely separate from these since, at their core, they use advanced AI and Deep learning techniques to create more realistic media with a higher chance of detection, making them far more dangerous and convincing.
Are deepfakes illegal?
Deepfakes by themselves are not illegal; however, like with any technology, abusing deepfakes to target individuals or organizations, extortion and blackmail purposes, etc., is illegal as it constitutes cyber fraud, identity theft, and several other punishable offenses.
Can deepfake be detected?
Yes, deepfakes can be detected either manually by looking for mismatches in the content (as detailed in prior sections of this article) as well as via advanced deepfake recognition tools that can flag deepfake-generated media.
How can you tell if a video is AI-generated?
While deepfakes are extremely convincing if created properly, they cannot emulate real media perfectly. A discerning user will be able to spot inconsistencies in deepfakes, such as video background blurring, desync between audio and video, and other tell-tale signs that a particular video or image is generated by AI. Refer to the previous sections in this article on how Deepfakes can be identified for more.
What is an example of a deep fake?
Perhaps the most notable example of deepfakes, which skyrocketed them into fame, is the 2018 deepfakes of former president Barrack Obama. In these, he is seen making a public service announcement, but in reality, the contents of what he is speaking about are entirely AI-generated. While his voice and mannerisms are similar to real life, the content itself was never actually spoken by Barrack Obama.
