Covert Communication and Image Authentication Algorithm Based on Adversarial Examples

• Published in: International Symposium on Communications and Information Technologies (Online) pp. 199 - 203
• Main Authors: Wu, Qiwen, Feng, Zijing, Huang, Yingkai, Zhong, Jiehui, Liu, Xiaolong
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27.09.2022
• Subjects: adversarial examples; Authentication; covert communication; Deep learning; image authentication; Image coding; Neural networks; Receivers; Steganography; Tensors
• ISSN: 2643-6175
• DOI: 10.1109/ISCIT55906.2022.9931289

The research on the application of adversarial examples mainly focuses on considering adversarial examples as a threat in the past. In order to make better use of the adversarial examples, rather than just taking it as the defects of neural network, a covert communication method based on adversarial examples is proposed in this paper. Using the characteristics that adversarial examples can carry information, we combine it with specific coding rules to develop a covert communication algorithm. Unlike the traditional steganography, the secret information is not contained in the communication content of the sender and the receiver itself in the proposed scheme. The mapping relationship between adversarial examples and secret information is hidden in the neural network model, so as to realize the hidden transmission of information and improve the concealment and security of communication. At the same time, the tamper identification is embedded in the tensor of adversarial output. When the image is modified during transmission, the tamper identification will also change, so that the image can be authenticated. Experiments show the feasibility of the algorithm and verify that it can completely extract secret information from the encrypted image adversarial example and authenticate the integrity of the image.