Improving security for image steganography using content-adaptive adversarial perturbations

Improving security for image steganography using content-adaptive adversarial perturbations

연구 분야: Strategies

학회: Applied Intelligence

Cover enhancement is an important adversarial steganography method used to enhance the security of image steganography, which utilizes adversarial perturbations to attack deep learning-based steganalyzers. However, current adversarial steganography methods tend to introduce unexpected and detectable artifacts without considering the characteristics of image contents. In this paper, content-adaptive adversarial steganography (CAAS) is proposed to enhance the security of image steganography by adaptively adding perturbations into cover images considering image contents with rich texture, where perturbations are generated by adversarial example generation methods, such as the fast gradient sign method. In CAAS, a hybrid texture descriptor is first designed to describe the texture regions by applying the improved local binary pattern based on multi-grained gradient information and the noise residual feature. Then, a segmentation method, namely simple linear iterative clustering, is used to divide the input image into several regions by leveraging local semantics. Finally, a weighted mask is constructed based on the hybrid texture descriptor and segmentation results, which can be used to determine optimal positions for assigning adversarial perturbations with different weights to generate adversarial cover images with better security. Extensive experiments are conducted to compare with other state-of-the-art methods to verify the superiority of the proposed method. Experimental results show that the proposed CAAS can improve security in image steganography and cause fewer detectable traces.