Thoughts From The Deepfakes Arms Race

Forbes

25-03-2025

It's now sort of common knowledge that we're headed into a new AI being able to create sophisticated fake content – images, videos, etc.
This has huge ramifications for our world – from celebrities and politicians to normal people, everyone is going to be affected. We're already seeing these technologies being used by bad faith factors in mass media, and we're seeing young people creating fake nudes of each other in ways that are harmful.
So what do you do? How do we push back against the harmful effects of a new threat?
I wanted to highlight some of the ways that both deepfake technologies and defensive systems are getting better over time. That might help us think about how to tackle this phenomenon.
The most sophisticated deepfake spotter tools can take a video and look for those subtle tells that show whether it's naturally produced or not.
This IEEE piece goes over some of the forensic markers that can be used, and points out that many of these tools are ultimately effective.
There's a human component, too, as in this piece where Louise Bruder is described as a 'super-recognizer' tasked with spotting deepfake content.
Here's more from a team presenting in a recent TED talk around a particular tool called Deep Detect:
Melat Ghebreselassie and Elon Raya are explaining the use of this technology and its benefit to the public.
Raya points out that you need something that is live, in his words, 'on the go', easily accessible, and versatile. Deep Detect, he says, works in real time on platforms like YouTube, which is where we need this type of work to be done.
On the other hand, former solutions lacked efficacy.
'They just crumbled,' he said of traditional tools, 'because they're trained in controlled environments.' New tools, he notes, are trained with messy data, so they are better at identifying deepfakes, in the wild, as it were.
In describing these systems, Ghebreselassie and Raya go over that interplay between two effective AI engines.
You could call it the 'generative adversarial network model,' which was a common way to describe this in the earlier days of machine learning, at the beginning of this decade.
In this situation, a generative model is producing things: in this case, deepfakes. The adversarial model is assessing them, and telling if they're deepfakes or not. The traditional idea is that the result goes back to the generative model to tweak or fine-tune its results. However, if the result goes to a human audience instead, it works against the deepfaker.
In other words, this can be used in both positive and negative ways with deepfakes. The adversarial engine can be detecting the deepfakes and pointing them out to the public, or it can be training the generative network to make even better deepfake content.
However, presumably, there's a limit to what deepfakes can accomplish. They may not be able to really mimic all of these small nuances of naturally produced content. For example, one way that the deepfake spotters work is to look at lighting and natural composition. In order to fool these tools, the deepfake generator would have to create all of the dynamics of natural lighting, down to the very last detail. The extent to which new models can do that is sort of unknown.
Then there is human involvement, as in the BBC piece that I mentioned. People are the gatekeepers in some of these situations, and they can use the AI in assistive ways
In terms of the technology that Raya and Ghebreselassie present, there's a sort of crowdsourcing model that Ghebreselassie calls a 'game-changer' for deepfake spotting.
In decentralized finance, a crowd of people forms a consensus, and that's the trust mechanism for the economic outcome. You can do that with deepfake detection, too. If 90% of people report it as a potential deepfake, you can take that into account as algorithms determine what's likely to be fake content.
'If 80% of users agree on a piece of content, it's added for retraining and flagged for our model to learn from,' Ghebreselassie explains. 'And we're using replay-based learning that allows the model to learn without forgetting the past pieces of content it's seen.'
'We use regularization methods such as elastic weight modulation,' Raya adds, 'so that the model doesn't forget key features learned, while also fine-tuning on new data.'
In addition, Ghebreselassie pointed out that the Deep Detect system has two 'superpowers': one is a multi-head attention mechanism, allowing it to focus on more than one thing at once. Another is the ability to normalize with NLP.
The system uses convolutional neural networks to extract features, and a visual transformer to perform deep evaluations.
So the news here is mixed. The good news is that the deepfake spotting tools are pretty elaborate and complex. They're pretty advanced. They use all of the above power to allow us to minimize the harm that deepfakes are causing. The bad news is that the deepfake generators are always evolving, too.
Feel free to drop a comment below if you have a thought on how this is going to impact our society in the years to come.