Hardness of Learning AES with Gradient-Based Methods.

We show the approximate pairwise orthogonality of a class of functions formed by a single AES output bit under the assumption that all of its round keys except the initial one are independent. This result implies the hardness of learning AES encryption (and decryption) with gradient-based methods. The proof relies on the Boas-Bellman type of inequality in inner-product spaces.