It could, but within a given neural network structure, the speed is going to pretty much be constant. (Barring optimizations such as eliminating zero weights/activations). There's a meta-level above this of trying to search or automatically determine a "good" NN structure that can accomplish the encryption & decryption. That too (determining an optimal NN structure for a problem) is a fascinating research question in its own right! :) In fact, it's one that Jeff Dean called out a while ago as one of the leading-edge questions for DNNs, IIRC.