Research Benchmark: Can NNs break cryptography schemes?

Create a table of substitution ciphers (https://en.wikipedia.org/wiki/Substitution_cipher)

Create a database of books. Encrypt all of the books with a substitution cipher. (Is there a corpus of free books available somewhere?)

Run the books through progressively deeper neural networks, training them to break the code.

Create a table where the rows are the substitutions and the columns are the neural depth.

Is there a trend?

Instead of substitutions, a similar benchmark could be made with other cryptographic algorithms.