AI that detects AI seems entirely plausible. And like all “anti” measures, is another arms race (and if I put my scifi hat on, is what may lead to AI self-awareness).
Sort of reminds me of a talk Valve gave about creating an anti-cheat for Counter-Strike using AI. When asked if they were worried about people using AI to create cheats to fool the AI, his answer was essentially that it was an arms-race won by the person with more data/processing power. That person would most likely always be Valve.
It's a nice sentiment, but there are popular and easy-to-find cheating projects (not sure if I can name them here) that are still widely used, these projects have been active for years, before that talk was made, and still active today. Based on youtube videos and comments it seems many users are still using these cheats with little issue. And afaik, the one I'm referring to (initials P.I.) doesn't use any machine learning at all.
In Greg Egan's _Permutation City_ this is mentioned in passing (an arms race between AI video call spam bots and AIs screening the calls for spam while impersonating the owner of the phone).
The anti-spam software loses because eventually having self-aware AI view spam calls 24/7 was considered torture and they weren't allowed to go that far.
Typically the AI to detect the AI is exactly what they use to train networks like Deep Fake in the first place. GANs are effectively just a local arms race between two machine learning networks.
Is it that difficult to create/retrofit an AV container format for cryptographically signed audio and video streams? Key management & revocation could be a pain, but it's something that consumer electronics companies like Apple could do: think of it as the MPEG LA, but with signature checking & non-repudiation.
Congratulations, you have now created a class of people who can forge audio and video streams at will, relatively cheaply - and an underclass of people who cannot and may not even be able to record genuine footage they wish to record. This is not a road you want to go down.
No - I just created the equivalent to https for video; the "underclass" can still create, share and play unsigned videos - those would get low-trust warnings[1] (as they should, just like there is an "underclass" with no cert for their site). This wouldn't take away anything from todays' tech, only adds attestation for person/org behind videos they would like to mark as "official".
1. (edit) It occurred to me that some people may wish to manage the public keys independent of (say, Apple) and they could distribute via keybase or key-signing parties, so they actually don't have to suffer low-trust warnings. Now that I think of it, instead of merely signing streams, they could be signed and encrypted using recipients PK for 1:1 transmissions. Obviously law enforcement won't be a fan
Law enforcement will just coerce the CA system you've suggested to secretly improperly issue certificates. Just as it does with the current CA system. Problem not solved - it's just hidden.
You conflate content trustworthyness with origin sureity but a CA system doesn't even provide that.
Being good at Photoshop is really difficult and producing good fakes is extremely time consuming. Today, even with a Hollywood budget, most such effects still look off. However, the industry has gotten much with actor enhancement for example generally going unnoticed.
Which I think is real issue, this stuff is becoming easier over time. AI could be the tipping point where eventually people just stop trusting images and video. But, that transition is gonna be difficult.
Good luck with that. The best product you'll come up with is some sort of snake oil. The whole point of GANs is that you can't really "detect" the synthesized components anymore. Not that this would/will stop people from claiming otherwise in the spirit of profitablity :-)
No, GANs train exactly one discriminator, jointly with the generator. There's no guarantee that you can't train another good discriminator out of band.
Furthermore, GAN discriminators are (as I understand it) often hobbled a bit to ensure that the generator can make progress on the loss function. An always-correct D doesn't provide a useful gradient.
GANs train by fooling AIs, not humans. Fooling humans is a side effect, not the primary thing trained for (mostly because that's cheaper of course). It's just that humans are in some ways different from AIs in terms of fooling.
Looking at the papers I must say I think the ability to fool humans is a scale problem, not a fundamental limitation. Already GAN produced images and sounds survive "normal" human scrutiny: if you have no reason to suspect foul play you won't see it. If you really go looking, you'll see it.
Snake oil-tier anti-ML might do the trick. One of the problems is the level of confidence people put into ML in the first, when a lot of times it's also snake oil-quality. Just being able to cast doubt on that again would help prevent a loss of healthy skepticism and critical thinking.
Better provenance is the way to guard against fakes.
For video and audio, I imagine a combination of hardware signing, perhaps with the camera itself living on an isolated, Secure Enclave-like chip, and sending hashes of (incoming images/video * deviceID * trustedTimestamp) to a blockchain or some other public distributed ledger. Getting the timestamp from a service that keeps its own record adds further security.
This obviously requires an internet connection, and would likely be useful mostly for news and government agencies, law enforcement. But if the culture is affected enough by deepfakes, I can imagine it becoming more ubiquitous. The parts are all there, it’s a question of utility.
Sure, that’s the reason for including the 3rd-party timestamp AND for keeping the device signing keys on a separate secure chip. The idea is to say “these images/sounds were recorded at this time by this device,” and to have that statement both registered publicly at the time of creation and backed by the reputation of the device maker.
It’s acknowledging that using AI to catch AI fakes is a fool’s errand, and relies instead on the premise that hashing a raw data stream is much faster than producing a good fake, and that a secure device key is secure. You’d need both for it to work, otherwise you can generate a deepfake beforehand and get the device to sign a fake stream. That may be easier to do than I think; this is not my area of expertise.
ah yes, if u listen carefully to the samples, you can always tell subtle things that make it seem a little off. Maybe if you look at the binary data very carefully, it would be easy to show HUMAN_AUTHENTIC or CREATED_BY_MACHINE and sell this service. Someone have a recording of something you never said? For $99.99 get it checked at AreYouHumanDotCom!
It may be that the next big business opportunity lies in creating 'anti-AI' technology just as it did with antiviruses in the 90's and 2000's