3 ms·
Touching off of this: We have access to the learned encryption function, f : plaintext -> ciphertext. Most importantly, f is differentiable. It has to be, sinc
by gcr 10y ago
Touching off of this: We have access to the learned encryption function, f : plaintext -> ciphertext.
Most importantly, f is differentiable. It has to be, since it was trained with gradient descent.
So if you want to decipher a ciphertext Y, then use backpropagation to find the X that minimizes distance(f(X), Y). You already know dF/dX (f is differentiable after all), so run this through your magic optimization solver and you'll get an X very close to the result.
Researchers use this "backproagation in input space" idea to recover the training set of a trained neural network, or to debug their model by finding inputs that activate certain neurons, for example. (This is the basic idea behind Deepdream from last year)
It's not even correct to call this "cryptography." The transformation isn't one-way.
- contravariant 10y agoCryptographical transformations aren't supposed to be one-way. Not if you know the key anyway. It would be pretty pointless to encrypt something in a way that can't ever be read.
- andorov 10y agoIt is useful in cases where you want to verify something is the same as a stored (encrypted) value - e.g. passwords. One-way encryption where the entire process including the equivalent of salt/pepper is publicly available and not reversible might have some interesting applications.
- pizza 10y agoNot that this is relevant to OP, but there are also non-differentiable descent (edit: NOT descent, but rather just general optimization) methods; https://optimization.mccormick.northwestern.edu/index.php/Nondifferentiable_Optimization#-Subgradient_Method https://optimization.mccormick.northwestern.edu/index.php/No...