Towards secured image steganography based on content-adaptive adversarial perturbation

Abstract

In this paper, a content-adaptive adversarial steganography is proposed to improve steganography security by adaptively adding perturbations to cover images while considering image contents with rich texture, where perturbations are generated using adversarial example generation methods like the fast gradient sign method. In this approach, a hybrid texture descriptor is initially created to describe texture regions by using better local binary patterns based on multi-grained gradient information and the noise residual feature to describe texture regions. The input image is then divided into different sections using local semantics using a segmentation approach called simple linear iterative clustering. Finally, using the hybrid texture descriptor and segmentation results, a weighted mask is created, which may be used to determine the best spots for applying adversarial perturbations of various weights to generate more secure adversarial cover images. Extensive experiments are carried out to compare the suggested method to existing state-of-the-art methods in order to prove its superiority. The experiments were conducted on BOSSbase ver. 1.01, which contains 10,000 grayscale 512*512 images. The images were cropped into four non-overlapping 256*256 images using ‘imcrop’ function in MATLAB R2018b. Consequently, a cropped BOSSbase dataset was constructed that contains 40,000 samples. Besides, we also evaluate the performance on another image dataset, namely BOWS2. Experiments show that the proposed model can increase image steganography security while causing less observable traces.