On the Practical Uses of Experimental Adversarial Neural Cryptography

Abstract

With the rise of generative adversarial networks (GANs), many areas have seen remarkable improvements, e.g., computer vision, natural language processing, and the medical field. Notably, cryptography has been fueled by GANs producing adversarial neural cryptography (ANC). However, in these five years, ANC has little documented experimentation and applications that can be used in the real world. This paper aims to perform experiments on ANC to verify if the current status of ANC is ready for practical implementations of symmetric-key encryption. In our investigation, we assess several entities in ANC during training, encryption, and decryption of an ANC model, including decryption accuracy analysis. Furthermore, we study the resources required for deployment using different quantization techniques to reduce the size of an ANC model and its impact on performance and decryption accuracy. Our study provides enough data for offering practical advice for using and implementing ANC models.