Automatic encryption schemes based on the neural networks: Analysis and discussions on the various adversarial models (short paper)

Modern cryptographic schemes have been focusing on protecting attacks from computational bounded adversaries. The various cryptographic primitives are designed concretely following some randomization design strategies, so that one of the goals is to make it hard for the attacker to distinguish between the real ciphers and the randomly distributed ones. Recently, Google Brain team proposed the idea to build cryptographic scheme automatically based on the neural network, and they claim that the scheme can defeat neural network adversaries. While it is a whole new direction, the security of the underlined scheme is remained unknown. In this paper, we investigate their basic statistical behavior from traditional cryptography’s point of view and extend their original scheme to discuss how the encryption protocol behave under a much more stronger adversary.