Steganogram removal using multidirectional diffusion in fourier domain while preserving perceptual image quality
Introduction
With the emergence of Internet as a most sought after mode of communication, the need for securing such communications has also become primary. Further, the proliferation of multimedia content and easy and instant sharing facility of images in social media apps have resulted in making images a more preferred data than text. Data espionage attempts within an organisation are using steganography and these image covers help them greatly to hide confidential data like organisation credentials, source code, defense plans, trade secrets, malware programs etc. easily as well as without creating any suspicion.
Steganalysis is the currently employed popular counter measure to evade against steganography [[9],[10],[12]]. However, the higher-order statistical and machine learning based steganography detection techniques are both computationally expensive as well as not universal (all types of steganography techniques). Recently, convolutional neural networks (CNNs) also have been adopted for image staganalysis [24]. New steganography techniques are released so frequently the steganalysers could not keep up the pace in getting exposed/trained to those signatures and hence suffer from high false alarms which limits their usefulness as a within- enterprise security solutions against steganography. The philosophy of stego destruction is different from the steganalysis process. It is a technique that attempts to destroy only the hidden information content within the digital medium without inflicting any harm to the cover image. A steganalyser with high false alarm rate imposes inconvenience to a genuine user when it misclassifies a clean image as a stego one. At the same time, the organisation cannot afford leakage of confidential information by lowering the threshold of the steganalysers. Accordingly, there is a need for a simple but effective, universal solution to destruct only the stego content at the same time preserving the visual quality of the images.
Previous works ([[1],[17],[18],[20]] suffer from one or more of the flaws:
- 1
They operate without an understanding of the steganography embedding algorithms. They blindly perform filtering and random overwriting of the images. The process might affect the picture quality of a clean/non-stego image.
- 2
They need some information about steganography scheme deployed, secret information size, and cover image which are mostly oblivious in general.
- 3
They affect the visual quality of the images considerably during their attempt to destruct the stego content.
Fawzi Al-Naim et al. [31] employed wavelet based denoising to destroy the stego content in the images. However a detailed investigation on spatial and transform domain steganography schemes and a test on large dataset needed to be implemented for proving the robustness of the approach.
Amritha et al. [2],[30] employed showering algorithms which involve radiometric and geometric transformations applied on the stego images generated by HUGO-BD, WOW, Synch and J-UNIWARD. They showed both a direct evaluation of stego removal as well as an evaluation using universal steganalysis. They handled texture images and non-textured images separately and employed GLCM features to characterize the image textures. Their method is effective on texture cover images. However for natural images, this method could not produce superior results, since their choices of filters like Median, Gaussian, Wiener filters operated well on texture images. They also suggest which filters are more suitable for which type of images – textured or non-textured.
Dahuin Jung et al. [28] proposed a deep neural network based solution that exploits sophisticated pixel distributions and edge regions of images and erases stego images generated by deep learning steganography schemes. This work is a kind of partly active steganalysis, though not a complete one, since it removes the secret image and purifies the stego image to look very much similar to the cover image. However, all the stego images contain images as the secret message, and the deep learning based steganography scheme takes cover image and secret images as inputs and performs data embedding while producing the stego image. The complete image characteristics is explored and exploited to facilitate the data embedding. So, analysis on diverse secret messages like text and other file types are not investigated. Further the steganography schemes that operate in spatial domain and the transform domain are not dealt separately. The trained model is capable of removing messages generated by deep steganography schemes like Deep Steganography [35] and invisible steganography via GAN (ISGAN) [34].
Corley et al. [29] developed a steganography purification system called as Deep Digital Steganography Purifier (DDSP), which employed a Generative Adversarial Network (GAN) that erased the steganographic content from cover images while preserving good perceptual image quality. However, the model has to be extended to cover media and secret object of various types, sizes, and color spaces. Further the model has to be more robust to make it ready for practical use, which can be done by training the GAN using a large dataset.
Despite of the remarkable works done in this area [[16],[23],[27]], no low-complexity technique that would systematically destroy the embedded messages in images has yet been proposed.
This paper proposes such a solution by modeling the stego embedding process as an additive noise model and removes the stego noise by an iterative diffusion process in the Fourier domain.
Section snippets
Steganography process
A grey scale image can be denoted by the set: where represents the pixel coordinates and is the corresponding intensity value. For a typical image when represented in integer or for a real value representation of the intensity .That image can also be denoted by the matrix where denotes the intensity of the pixel at the position.
Steganography embedding is modelled as a stego noise
Preparation of test images and schemes
We perform the experiments aimed to evaluate: 1) generalization – ability to remove stego-noise created using different embedding methods, 2) performance – ability to maintain image quality, and 3) robustness - preserve the fine details as well as edge regions of the image.
We use the following stego schemes: Cox et al. [7], Digimarc [8], DWT [13], JSteg [[14]], PGS [[15]], Steganos [[22]], S-Tools [[6]], YASS [21]
Though watermarking schemes are not used for secret communication, we chose them
Statistical analysis
To compare the performance of different stegoschemes and to assess the statistical significance of the results, we adopted the non-parametric Friedman test and post-hoc Nemenyi test for the compared methods across multiple images. The results of the Nemenyi test regarding PSNR and SMC (see Fig. 7, Fig. 8) show that the differences between the stegoschemes are statistically significant (p < 0.001). Critical Difference (CD) shows the smallest difference in mean ranks, where the difference is not
Discussion
Our main results are as follows:
- 1)
The stego content destruction ability offered by our method is very good for any steganographic scheme in general, with small loss of visual quality of the original image content. The novel multi-directional diffusion process in the Fourier domain facilitates scrubbing of the stego noise content while preserving the original image content.
- 2)
The PSNR and SSIM values prove that the proposed system is a simple yet efficient mechanism to remove the stego content. The
Conclusions
In this paper, a novel and efficient method for destructing any stego information hidden (steganograms) inside cover images by any of the steganographic technique is proposed, which also sustains the perceptual quality of the images as well. Iterative multi-directional diffusion process in the Fourier domain is executed, which corrupts the stego bits embedded in the image, and this is repeated until the visual quality of the image does not drop below the desired threshold. This proposed method
Declaration of Competing Interest
The authors declare no conflict of interest.
Acknowledgments
The authors are grateful to VIT management for providing the Research Seed Grant (AY 2019-20) to execute this project/work.
References (35)
- et al.
A review of image steganalysis techniques for digital forensics
J. Inform. Security Appl.
(2018) - et al.
FAUST: Forensic artifacts of uninstalled steganography tools
Digital Invest.
(2009) - et al.
Optimal image steganography content destruction techniques
- et al.
Anti-forensic approach to remove stego content from images and videos
J. Cyber Security Mobility
(2019) - et al.
A survey on different feature extraction and classification techniques used in image steganalysis
J. Inform. Security
(2017) - et al.
A method for automatic identification of signatures of steganography software
IEEE Trans. Inf. Forensics Secur.
(2010) - Blasco, J., Hernandez-Castro, J., de Fuentes, J., Ramos, B., 2012. A framework for avoiding steganography usage over...
- Brown, A. S-tools version 4.0. [Online]. Available:...
- et al.
Secure spread spectrum watermarking for multimedia
IEEE Trans. Image Process.
(1997) - Digimarc Corp., 2015. PictureMarc, Embed Watermark, v...
Dimensionality reduction for classification of blind steganalysis
J. Software Eng.
Blind statistical steganalysis of additive steganography using wavelet higher order statistics
Lecture Notes in Computer Science
Wavelet based watermarking method for digital images using the human visual system
Electron. Lett.
Protection of mammograms using blind steganography and watermarking
Cited by (0)
Editor: Maria De Marsico