关键词: 抗泄露, 生成对抗网络, 批规格化, 全连接神经网络, 卷积神经网络

Abstract:

Generative Adversarial Networks（GANs） is a kind of deep learning model. Through adversarial training with discriminative model, a gradually improved generative model can be obtained to generate data which are difficult to distinguish between true and false. And using GANs to realize encryption algorithm is a new research direction. Firstly, under the 16 bit key symmetric encryption scheme, this paper tests the key resilient encryption communication of the basic encryption communication model built by Abadi et al., and finds the possibility of realizing the key resilient encryption communication by using GANs. Then, the neural network model of both parties and the adversary is improved by modifying the activation function of the network, and obtains the encryption algorithm model in the case of 3 bit key leakage. The communication stability can be improved by increasing the complexity of the decrypter and the adversary model. Then, adding batch normalization in the model can further improve the ability of key resilient encryption communication. Finally, in the case of 8 bit leakage, it can ensure the normal communication between the two sides of the communication and the adversary cannot obtain the secret information. This paper provides a new solution to the problem of key resilient encryption communication and proves the feasibility of the solution through experiments.

Key words: key resilient, generative adversarial networks, batch normalization, fully connected neural networks, convolutional neural networks