If you can write a function to quantify those errors, then you can use it to train the neural networks that produce the deepfakes. The better your error function, the better the deepfakes.
It sounds very much like complexity theory, where you definitely have problems that are easier to verify than solve (see NP hard).
(Otherwise I could do stuff like use a hash function to train a NN to compute input for hashes , for example, no?)
In theory, yes - you can. The training function could return a score - say the number of matching digits in the hash - and the NN will in theory learn what inputs produce the better output. If somehow, there is a weakness in the hash algorithm, it could stumble onto it - allowing it to get better at producing the right input to get the required output.
The simpler it is to deterministically manipulate, the easier it should be for a NN to learn to manipulate - even of the function is just returning a boolean or a "rating between 1 and 10". So yea, good hash functions are unlikely to be learned and solved, but photos and videos aren't designed to be hash functions.
(All that said, I'm willing to be bet the time and computing power you'd need to pour into this NN to break a good hash function is likely more than has ever existed in the sum of all past time + computing resources ;))
I think you mean NP. NP hard includes problems that are not in NP, such as the halting problem.
I'm not aware of any proof that photorealistic 3D rendering is in NP. If it is not, then verifying 3D cannot be said to be easy.
[1] https://en.wikipedia.org/wiki/P_versus_NP_problem#Reasons_to...
The main argument from your link is nobody has found an efficient algorithm for any of the 3,000+ studied problems after all this time.
And yet it’s also true that in the same span of time, nobody has been able to prove the that it’s not, either. No matter how hard they have tried.
“probably not” is just an opinion. I’ll wait for the formal proof. Till then it simply is not known whether or not.
if there was some score that would look at lighting etc, you probably could just tack it on to one of the loss functions somewhere and expect to see improvement.
maybe not enough to beat the system; or perhaps it would increase the lighting score but make the image obviously unrealistic in other ways we haven’t thought of.
basically at this point you would have a lot of if-else statements, like a "Expert System" AI in the 80s.
so much so that it would be better to use a NN to replace the loss/error/verification function instead of coding it by hand?
1) computationally intractable to forge. (e.g. requires simulating trillions of photons)
2) computationally tractable to verify
Sadly, my immediate gut intuition is that there is not such a problem, for a variety of reasons; but hopefully I'm wrong!
You can't solve the travelling salesman problem in polynomial time just by throwing machine learning at it...
On a mathematical level you can prove that machine learning will do a bad job for certain things, and this is what the parent talked about.
https://www.cc.gatech.edu/~lsong/papers/arxiv_rl_combopt.pdf
It does at least respectably well against the Concorde TSP solver although (unsurprisingly) doesn't beat it. Note that this also isn't a neural network that just maps a graph to a TSP solution; instead, the neural network function is a heuristic that guides the greedy construction of a solution one vertex at a time.
For instance one fundamental of forensic accounting is Benford's law [1]. Even in peripherally related with artistry it stands out. one of my never off the ground webcomic attempts with digital art I tried making backgrounds with perspective and I calculated the pixel spacing completely evenly - it stood out as unnaturally regular compared to doing it with just a straight-edge and a ruler on a cheap Wacom tablet.