Covert Communication via Modulating Soft Label of Neural Network

Abstract

AbstractIn this paper, we present a novel method to covert communication based on a host deep neural network (DNN) itself, which is totally different from many traditional works that embed secret data into a digital image since: 1) there has no direct image transmission between the data hider and the data receiver, 2) there has no modification to the image content, and 3) the presence of the covert communication can be concealed by the large number of ordinary queries. The main idea is embedding secret data into the soft output of the host DNN by modifying the soft label according to the secret data and further fine-tuning the host DNN with the modified soft label together with its input. When the fine-tuned DNN is put into use, the data receiver can fully retrieve the secret data from the soft output of the DNN by uploading the corresponding image to the DNN. Experimental results have shown that, the secret data can be successfully embedded and extracted. And, it does not impair the performance of the original task, which has demonstrated the superiority and applicability. Moreover, the Kullback-Leibler divergence between the soft output produced by the embedded DNN and that produced by the original DNN is quite low, which can ensure the security.KeywordsCovert communicationSteganographyInformation hidingDeep learningFine-tuningDeep neural networks