Enhancing JPEG Steganography using Iterative Adversarial Examples

Abstract

Convolutional Neural Networks (CNN) based methods have significantly improved the performance of image steganalysis compared with conventional ones based on hand-crafted features. However, many existing literatures on computer vision have pointed out that those effective CNN-based methods can be easily fooled by adversarial examples [1]. In this paper, we propose a novel steganography framework based on adversarial example in an iterative manner. The proposed framework first starts from an existing embedding cost, such as J-UNIWARD [2] in this work, and then updates the cost iteratively based on adversarial examples derived from a series of steganalytic networks until achieving satisfactory results. We carefully analyze two important factors that would affect the security performance of the proposed framewrork, i.e. the percentage of selected gradients with larger amplitude and the adversarial intensity to modify embedding cost. The experimental results evaluated on three modern steganalytic models, including GFR, SCA-GFR and SRNet, show that the proposed framework is very promising to enhance the security performances of JPEG steganography.

HCISNet: Higher‐capacity invisible image steganographic network

“…Under the wave of deep learning, some pioneering approaches leveraged deep networks to hide and extract information with robustness and information capacity. Mo et al [18] proposed a generic network against embedding to improve the security and effectiveness of JPEG steganography, which can resist detection by steganalysers. Liu et al [19] developed a two-stage deep learning network for blind watermarking, using the encoder-decoder network to embed the copyright information, which can achieve resistance to high-intensity noise and enhance network stability.…”

“…Mo et al. [18] proposed a generic network against embedding to improve the security and effectiveness of JPEG steganography, which can resist detection by steganalysers. Liu et al.…”

Enhancing Image Steganography Via Stego Generation And Selection

“…This work was supported in part by the National Science Foundation of China (61972430), in part by the Natural Science Foundation of Guangdong (2019A1515011549) methods are symmetric, meaning that the costs of two directions (i.e., ±1) are exactly the same for every embedding unit. Recently, some asymmetric methods such as [6,7,8,9] and CNN-based methods (e.g., ADV-EMB [10], MinMax [11] [12], JS-IAE [13]) have been proposed. These methods take the direction of modifications into account and further improve security.…”

“…For instance, ADV-EMB [10] divides all D-CT coefficients into two non-overlapping parts: one for traditional embedding; the other for adversarial embedding. JS-IAE [13] updates existing embedding cost iteratively based on adversarial examples derived from a series of pretrained networks. In addition, the method [14] improves existing symmetric steganography methods by constructing enhance covers.…”

“…In this paper, we propose a new method to enhance existing steganogrpahic methods via stego generation and selection. The proposed method is inspired by some steganographic methods based on adversarial example (e.g., [10] and [13]) and our previous work [15] based on stego post-processing. In our method, we firstly train a steganalytic network based on the steganography to be enhanced, and then adjust a random part of embedding costs according to the magnitudes of original costs and the corresponding gradients from the pretrained network via back propagation.…”