The White House wants to ‘cryptographically verify’ videos of Joe Biden so viewers don’t mistake them for AI deepfakes::Biden’s AI advisor Ben Buchanan said a method of clearly verifying White House releases is “in the works.”
The White House wants to ‘cryptographically verify’ videos of Joe Biden so viewers don’t mistake them for AI deepfakes::Biden’s AI advisor Ben Buchanan said a method of clearly verifying White House releases is “in the works.”
Depending on the implementation, there are two cryptographic functions that might be used (perhaps in conjunction):
Cryptographic hash: An arbitrary amount of data (like a video file) is used to create a “hash”—a shorter, (effectively) unique text string. Anyone can run the file through the same function to see if it produces the same hash; if even a single bit of the file is changed, the hash will be completely different and you’ll know the data was altered.
Public key cryptography: A pair of keys are created, one of which can only encrypt data (but can’t decrypt its own output), and the other, “public” key can only decrypt data that was encrypted by the first key. Users (like the White House) can post their public key on their website; then if a subsequent message purporting to come from that user can be decrypted using their public key, it proves it came from them.
A note on this. There are other videos that will hash to the same value as a legitimate video. Finding one that is coherent is extraordinarily difficult. Maybe a state actor could do it?
But for practical purposes, it’ll do the job. Hell, if a doctored video with the same hash comes out, the White House could just say no, we punished this one, and that alone would be remarkable.
You’d need to find one that was not just coherent, but that looked convincing and differed in a way that was useful to you—and that likely wouldn’t be guaranteed, even theoretically.
Even for a 4096 bit hash (which isn’t used afaik, usually only 1024 bit is used (but this could be outdated)), you only need to change 4096 bits on average. Even for a still 1080p image, that’s 1920x1080 pixels. If you change the least significant bit of each color channel, you get 6,220,800 bits you can change within anyone noticing. That means on average there are 1,518 identical-looking variations of any image with a given 4096 bit hash, on average. This goes down a lot when you factor in compression: those least significant bits aren’t going to stay the same. But using a video brings it up by orders of magnitude: rather than one image, you can tweak colors in every frame The difficulty doesn’t come from the existence, it comes because you need to check 2⁵¹² = 10¹⁵⁴ different images to guarantee you’ll find a match. Hash functions are designed to take a while to compute, so you’d have to run a supercomputer for an extremely long time to brute force a hash collision
This concept is known as ‘collision’ in cryptography. While technically true for weaker key sizes, there are entire fields of mathematics dedicated to probably ensuring collisions are cosmically unlikely. MD5 and SHA-1 have a small enough key space for collisions to be intentionally generated in a reasonable timeframe, which is why they have been deprecated for several years.
To my knowledge, SHA-2 with sufficiently large key size (2048) is still okay within the scope of modern computing, but beyond that, you’ll want to use Dilithium or Kyber CRYSTALS for quantum resistance.