State of AI 2024: Why deepfake detectors are failing (and we got it)
Last year there was a little bit of a hype around deepfake detectors: due to the large amount of genAI images a variety of products emerged that claimed significant accuracy in detecting fake images.?
Well, now even the media realised that these detectors aren’t doing a great job.
Lately even the Washington Post wrote extensively about it https://www.washingtonpost.com/technology/interactive/2024/ai-detection-tools-accuracy-deepfakes-election-2024/
Why are these tools failing? You may dare to ask.
A simple, jargon-free answer : detectors use the same ‘level’ of technology (or even sometimes ‘inferior’) i.e. diffusion based or alike (to date), as the (fake) generators. In this way, the detector will never be able to be ahead of the generators hence resulting in mediocre results.
And, sadly, there is no way out of it. Generators will always have access to the latest technologies hence detectors are playing a catch up game.
Then, should we outright stop using detectors? No, not really. I am saying we should not ‘rely’ purely on them.
I give an example.?
Antivirus on your pc: does it catch all the virus or malicious software? No, it’s role today is to do the ‘basic’ checks in case you caught something really easy to detect. For your security you should have 2FA authentications, strong password, cryptographic databases etc.
Using the analogy, the role of deepfake detectors is akin to the antivirus: it will be embedded in routine, basic checks but we should not expect to really make us safe.
And yes,.. we got it in January in our State of AI 2024 (trend prediction 3).
And it is happening in front of our eyes.
The market will develop from here and more security measures will be introduced to counteract the deepfakes (we suggested watermark/signature products or even blockchain alike could work).
I hope it helps.
#ai #artificialintelligence #business #technology #cybersecurity #innovation
—---------------------------