Quantum-secure encryption using autoencoding?

I'm spending spare time researching quantum-secure encryption. My current effort involves training a neural network on a dataset. The result of the trained net is N values. Using an autoencoder architecture it is possible to "set" the N values (the encryption key?) to "reverse generate" the training data. If the training data is a well-separated set of QR codes (the current idea, likely to change) it seems that the autoencoder will be able to reproduce one, given the corresponding N-value. QR codes are naturally error-correcting so even a near-miss result can be corrected. If the QR codes are randomly generated they can be used to encrypt data. The "key" for the encryption is both the NN architecture and the set of weights. Given the NN without the weights the attacker knows nothing and cannot recreate the NN. Given the weights but not the architecture (which would include thinks like dropout or a "special" layer) the attacker knows nothing. This generates a large space of essentially random data that can be "recreated" by the autoencoder to use as a one-time pad. I don't see how a quantum computer can decode this. Grover's algorithm can choose "one of many" but I can't figure out how to, nor what to, encode in the quantum bits that might be input to Grover. So it seems to me that an autoencoder architecture can be shown to be quantum secure.