Deepfakes have become a major challenge in the digital world in recent years. Image and video manipulation is nothing new – after all, the term “photoshopped” has long since entered the vernacular. However, the simplicity and speed with which deepfakes can be created takes this technology to a new level. While image editing software often requires manual skill and a lot of time, deepfake tools allow even non-experts to create deceptively realistic manipulations. This is facilitated by the democratization of technology: What used to be the preserve of experts with expensive hardware is now available to the general public thanks to open source software and tutorials.

But let’s start from the beginning

The history of deepfake technology began in the early 2010s with the further development of neural networks, in particular generative adversarial networks (GANs). However, deepfakes owe their name to a Reddit user who published the first videos in 2017 in which he inserted the faces of celebrities into other content. Since then, the technology has developed rapidly and is now also available to hobbyists through freely available software. With the rapid advancement of artificial intelligence (which goes hand in hand with increasingly powerful hardware), the question inevitably arises:

Will we still be able to recognize deepfakes in the future?

This question is by no means new, and many people have already asked it. In 2019, Hao Li – technology pioneer and professor at the University of Southern California – claimed that we would no longer be able to reliably detect deepfakes in six to twelve months. In 2024, we know that this prediction has not come true. Nevertheless, this question is certainly justified.

While the technology for creating deepfakes is becoming increasingly sophisticated, the methods for detecting them often lag behind. The reason for this lies in the nature of the technology itself: Deepfakes are based on generative models that learn through large amounts of data to create ever more realistic content. Recognition algorithms, in turn, must continuously adapt to these advances. A technological arms race in which one side tries to create manipulations that are as credible as possible, while the other side looks for ways to expose them. The principle is similar to the battle between computer viruses and anti-virus software. Defense software is always reactive.

Recognizing deepfakes requires training with the deepfakes that need to be recognized.

What methods are currently available to detect deepfakes?

Machine learning methods for deepfake detection

AI-supported methods are currently being used to identify deepfakes. These systems analyze visual and auditory characteristics to detect inconsistencies. For example, unnatural movement patterns or inconsistent light reflections can indicate manipulated material. These analyses now go into great detail, right down to blood flow analysis. However, as already mentioned, these methods are in a constant race against the increasingly sophisticated techniques for creating deepfakes.

Heatmap of a video conference image analyzed with DeepDetectAI

Cryptographic approaches as a solution?

The approach of using cryptographic methods for deepfake detection is promising and is increasingly being discussed in research and development.

This is similar to the case with certificates for websites

Cameras or other recording devices could insert a digital signature during the creation of images or videos. This signature would confirm the origin of the content and make subsequent manipulation impossible, as any changes would invalidate the signature.

This could be particularly helpful for journalism, law enforcement or social media platforms. For example, reporters could prove that a video was actually recorded on location. In court proceedings, signed videos could serve as reliable evidence. And platforms such as X (formerly Twitter) or YouTube could give preference to signed content and label it accordingly.

However, there are currently no established and internationally recognized certificate authorities (CAs) that have been developed specifically for the authentication of media content. Nevertheless, promising approaches, technologies and standards are emerging that could potentially develop into a kind of “certification authority for digital content”.

Existing approaches and initiatives

Some initiatives and technologies already aim to ensure the authenticity and integrity of digital media:

• Coalition for Content Provenance and Authenticity (C2PA):
  The C2PA, supported by Adobe, Microsoft, X(Twitter) and the BBC, is developing standards for authenticating digital content. Its aim is to provide media content with metadata containing information about the origin and edits. This metadata is embedded directly in the file and can be generated during recording.
• Adobe Content Authenticity Initiative (CAI):
  With CAI, Adobe wants to guarantee the authenticity of digital content. Cameras and image editing software should add metadata that shows who created and edited an image or video.
• Truepic:
  This company focuses on the verification of visual content. Truepic offers a platform that verifies images and videos directly at the source and adds cryptographic signatures to guarantee their authenticity.
• Project Origin:
  This initiative, supported by the BBC and Microsoft, aims to combat disinformation by authenticating news content. Similar to the C2PA, mechanisms are also used here to make the origin and editing of content traceable.
• Blockchain-based verification services:
  Work is also currently underway on blockchain solutions that store and verify metadata for authenticating digital content. The blockchain serves as a decentralized register that makes manipulation virtually impossible.

Despite the potential, there are also several challenges that need to be overcome before cryptographic processes can be widely used:

1. Standardization and acceptance: For digital signatures for media content to be effective, device manufacturers, platforms and governments would have to cooperate worldwide. Without international standardization, the benefits will remain limited.
2. Hardware requirements: The integration of secure chips, such as a Trusted Platform Module (TPM), into recording devices could increase manufacturing costs. Older devices without a signature function would no longer be considered trustworthy.
3. Manipulation of signature mechanisms: The security of such systems depends heavily on the integrity of the private keys. If these are compromised, attackers could create forged signatures.
4. Decentralization of control: The question of who controls the certification authorities (CAs) and manages the signatures is critical. A centralized system could create monopolistic structures and thus itself become vulnerable.
5. Dealing with unsigned content: Existing media content or content from older devices will not contain signatures. Automatically classifying them as potentially manipulated harbors the risk of misjudgments.

Although cryptographic approaches are promising, there is currently a lack of standardized solutions and broad acceptance.

Conclusion:

Technology – both artificial intelligence and hardware – will continue to develop at a rapid pace. We have reached a point where algorithms are increasingly learning independently and optimizing themselves. This presents us with major challenges, of which deepfakes and disinformation are just a few examples.

So, can we still recognize deepfakes in the future?
The answer is twofold: Yes, but it’s getting harder and harder. Advances in deepfake technology, particularly through the use of generative AI, may mean that some fakes are barely recognizable as such. Nevertheless, there is hope. Advances in deepfake detection, such as the combination of image, audio and other parameters, will also improve the possibilities for verification. Cryptographic approaches and digital signatures also offer the potential to detect deepfakes at an early stage or make it more difficult to create them. In the long term, however, a combination of technology, regulation and human judgment will be needed to reliably detect deepfakes.

The most likely method will be a hybrid solution: automated detection systems such as those being developed at DeepDetectAI, supported by cryptographic technologies and human expertise.