Optimization and Security Analysis of Adversarial Neural Cryptography Applications

Compared with traditional cryptographic algorithms, neural network encryption algorithm is more aggressive, which is one of the hot issues in the field of intelligent cryptanalysis. In this paper, a simple improvement is made to the generative adversarial network encryption algorithm model established by GoogleBrain. This paper tests the encryption of files by trained and saved encryptors and decrypters. The final model reaches the point that the Bob decryptor decrypted with 100% accuracy with a probability of more than 80%, while the attacker’s Eve decryptor decrypts with an accuracy of about 50% all the time, which is no different from a complete guess. Then this paper analyzes the principle that the neural network encryption algorithm can realize anti-attack. In addition, for the ciphertext encrypted by Alice encrypter, this paper conducted randomness test on the ciphertext and compared it with the ciphertext randomness encrypted by AES, 3DES and Blowfish algorithm. It was found that although the encryption algorithm model generated by adversarial network training has stronger anti-attack resistance due to the lack of replacement structure, etc. However, the randomness of its ciphertext is very poor and even shows a certain periodicity, so the future research direction of neural network encryption algorithm needs to improve the randomness of its ciphertext.