Elsevier

Signal Processing

Volume 200, November 2022, 108668
Signal Processing

Short communication
Adversarial robust image steganography against lossy JPEG compression

https://doi.org/10.1016/j.sigpro.2022.108668Get rights and content

Highlights

  • This is the first work that introduces adversarial mechanism into robust steganography to improve the security of steganography. Adversarial dither modulation algorithm is proposed to guide the dither directions via the signs and magnitudes of cover gradients calculated from pre-trained steganalyzer.

  • According to “Matching Robust” or “Upward Robust” scene, we select proper embedding elements to balance the robustness against JPEG compression and steganography security.

  • Extensive experiments show that the proposed adversarial robust steganography could significantly improve the steganography security and maintain strong robustness.

Abstract

Traditional image steganography assumes that the transmission channel is lossless, thus it can not assure the correct extraction of hidden secret messages in real social networks. Recently, some robust image steganographic methods have been proposed to resist lossy JPEG compression. In this paper, we first introduce adversarial mechanism into robust image steganography to enhance steganography security. Instead of modifying embedding costs in traditional adversarial embedding steganography, we propose adversarial dither modulation algorithm according to the cover gradients calculated from a pre-trained CNN (Convolutional Neural Networks)-based steganalyzer. This algorithm determines the dither directions via the signs and magnitudes of the cover gradients. In addition, we carefully select proper embedding elements according to the prior information of the transmission channel to obtain a good tradeoff between “robustness” and “security” for robust steganography. Experimental results show that the proposed method can achieve significant improvements on security while preserving very good robustness compared with related work.

Introduction

Image steganography is the technology of hiding secret messages within digital images in an imperceptible manner. To enhance the security of steganography, modern steganographic methods are designed under the framework of distortion minimization [1], and they are mainly focused on the design of embedding costs with the help of STC (Syndrome Trellis Codes) [2]. Up to now, many effective JPEG steganographic methods have been proposed, such as J-UNIWARD [3], UED [4], UERD [5] and J-MiPOD [6], [7]. Note that all above steganographic methods [3], [4], [5], [6], [7] are additive steganographic methods which the costs of +1 and 1 are symmetric. To further improve the security of existing steganography, several side-information steganography [8], [9], [10], [11] and adversarial embedding methods [12], [13], [14], [15] have been proposed. These non-additive steganographic methods modify the symmetric costs to asymmetric ones according to some additional information (i.e., quantization errors for side-information steganography and gradients obtained by the targeted steganalyzer for adversarial embedding steganography).

Traditional steganographic methods assume that the transmission channel is lossless. In real social networks, however, some lossy image processing, such as JPEG compression, scaling, and cropping, may be performed on those stego images so that the hidden secret messages can not be extracted correctly. To deal with this problem, robust steganography has been proposed recently. Different from traditional steganography, robust steganography should be robust against some lossy image processing to assure the correct extraction of secret messages from the resulting stegos undergone some lossy processing. At the same time, those stegos should not be detected easily by the steganalytic methods. Thus, both “robustness” and “security” should be simultaneously considered in robust steganography. In current social networks, such as Facebook, lossy JPEG compression is the most popular image processing due to limited bandwidth and storage. Thus, existing robust steganography mainly focuses on resisting the lossy JPEG compression channel.

According to the relationship between the quality factor of cover (i.e., QFcover) and the quality factor of the JPEG compression channel (i.e., QFchannel), existing robust steganographic methods are divided into two categories in [16]: “Matching Robust” (i.e., QFchannel=QFcover) [17], [18], [19], [20] and “Upward Robust” (i.e., QFchannelQFcover) [16], [21], [22], [23], [24]. For instance, methods [17], [18], [19] proposed that repetitive re-compression on the cover image is useful for enhancing robustness for “Matching Robust”. DMMR (Dither Modulation and Modification with Re-compression based robust steganography) [20] used double-checking STC-RS coding strategy and re-compression operation to further improve the robustness for “Matching Robust”. In “Upward Robust” scene, method [21] constructed compression-resistant domain based on the relative relationship of DCT coefficients in the same frequency of four adjacent DCT blocks, and proposed RS-STC coding strategy via combining RS codes and STC. Thereafter, DMAS (Dither Modulation-based robust Adaptive Steganography) [22] utilized dither modulation algorithm to construct robust embedding domain to improve the robustness and combined side-information steganography to enhance the security. E-DMAS (Enhancing Dither Modulation based robust Adaptive Steganography) [24] proposed STC-CRC coding strategy which combined CRC (Cyclical Redundancy Check) codes and STC to further improve the robustness and security compared with method [22]. GMAS (Generalized dither Modulation based robust Adaptive Steganography) [16] utilized estimated side-information steganography and ternary STC, and expanded the embedding domain to the lower frequency within the DCT blocks compared with DMAS. Method [25] improved the robustness of robust steganography via minimizing channel error rate. Method [26] proposed a simulated repetitive compression network to resist repetitive compression.

Up to now, most robust steganographic methods focus on image pre-processing and/or embedding domain selection to assure the robustness. The recent literature has shown that some modern methods can achieve good robustness against JPEG compression, especially for “Matching Robust”. However, the security of steganography is still far from satisfactory. In this paper, therefore, we propose an adversarial robust steganography in order to enhance the security of robust steganography. Inspired by some adversarial embedding methods [12], [13], [15], we first introduce adversarial mechanism into robust steganography. Different from adversarial embedding methods in traditional steganography which modify the original symmetric embedding costs to asymmetric, we propose an adversarial dither modulation algorithm according to the cover gradients calculated from a pre-trained CNN-based steganalyzer. The main idea of the proposed method is the design of adversarial modulation rule to determine the dither directions according to cover gradients signs and magnitudes. In addition, we carefully select embedding elements according to JPEG robust scenes (i.e., “Upward Robust” and “Matching Robust”) to obtain a good tradeoff between “robustness” and “security” based on experiments and analysis. Experiment results show that the proposed method can achieve significant security improvements while preserving very good robustness compared with related work. Overall, the main contributions of this paper are summarized as follows.

  • This is the first work that introduces adversarial mechanism into robust steganography. We propose adversarial dither modulation algorithm to determine the dither directions via the signs and magnitudes of cover gradients calculated from a pre-trained CNN-based steganalyzer.

  • According to “Matching Robust” or “Upward Robust” scene, we carefully select embedding elements to obtain a good tradeoff between “robustness against JPEG compression” and “steganography security”.

  • We provide extensive experiments to show that the proposed method is very promising to enhance the steganography security while maintaining very good robustness simultaneously.

The rest of this paper is as follows. Section 2 gives the preliminaries. Section 3 describes the proposed adversarial robust steganography. Section 4 shows experimental results and discussions. Finally, the concluding remarks of this paper and future work are given in Section 5.

Section snippets

Notations

In this paper, boldface symbols stand for matrices or vectors, and handwriting symbols stand for sets. Specially, the quantized DCT coefficients of cover JPEG image with size n1×n2 is denoted as X, and the corresponding stego is denoted as Y. The corresponding de-quantized DCT coefficients are denoted as Xdeq and Ydeq. In addition, the symbol J1(X) denotes the decompressed spatial image using quantized DCT coefficients X.

Dither modulation algorithm

Dither modulation algorithm is widely used in modern robust steganography

Proposed method

As illustrated in Fig 2, the proposed robust steganography includes two steps, that is, steganalyzer pre-training and adversarial stego generation. We will firstly describe them in the following two subsections, and then show the extracting procedure of the proposed method in the third subsection.

Experiment results and analysis

In our experiments, 20,000 grayscale images of size 512×512 are taken from BOSSBase [27] and BOWS2 [30]. As in [31] and [32], all the images are resized into 256×256 using “imresize” in Matlab with default settings and then JPEG compressed with quality factor 65 and 75 separately. For each cover set with a given quality factor, we divide it into three non-overlapping parts, that is, 14,000 images (10,000 from BOWS2 and 4,000 randomly selected from BOSSBase) for training, 1,000 images from

Conclusion

In this paper, we propose an adversarial robust steganography against lossy JPEG compression. To our best knowledgement, this is the first work that introducing adversarial mechanism into robust steganography. The proposed adversarial dither modulation algorithm designs adversarial modulation rule to determine the dither directions via the signs and magnitudes of cover gradients calculated from a pre-trained CNN-based steganalyzer. In addition, we carefully select embedding elements based on

CRediT authorship contribution statement

Minglin Liu: Conceptualization, Methodology, Software, Validation, Formal analysis, Writing – original draft. Hangyu Fan: Investigation, Writing – review & editing, Funding acquisition. Kangkang Wei: Data curation, Writing – review & editing. Weiqi Luo: Resources, Writing – review & editing, Funding acquisition, Project administration, Supervision. Wei Lu: Resources, Writing – review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

This work is supported by the National Science Foundation of China (U2001202, 61972430), the Natural Science Foundation of Guangdong (2019A1515011549), and the Alibaba Group through Alibaba Innovative Research (AIR) Program.

References (37)

  • R. Cogranne et al.

    Efficient steganography in jpeg images by minimizing performance of optimal detector

    IEEE Trans. Inf. Forensics Secur.

    (2021)
  • T. Denemark et al.

    Side-informed steganography with additive distortion

    2015 IEEE International Workshop on Information Forensics and Security, WIFS 2015, Roma, Italy, November 16–19, 2015

    (2015)
  • T. Denemark et al.

    Model based steganography with precover

    Media Watermarking, Security, and Forensics 2017, Burlingame, CA, USA, 29 January 2017 - 2 February 2017

    (2017)
  • W. Li et al.

    JPEG steganography with estimated side-information

    IEEE Trans. Circuits Syst. Video Technol.

    (2020)
  • Z. Wang et al.

    On improving distortion functions for JPEG steganography

    IEEE Access

    (2018)
  • W. Tang et al.

    CNN-based adversarial embedding for image steganography

    IEEE Trans. Inf. Forensics Secur.

    (2019)
  • S. Bernard et al.

    Explicit optimization of min max steganographic game

    IEEE Trans. Inf. Forensics Secur.

    (2021)
  • T. Song et al.

    Enhancing image steganography via stego generation and selection

    IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2021, Toronto, ON, Canada, June 6–11, 2021

    (2021)
  • Cited by (2)

    View full text