Image Steganography And Steganalysis.pdf
Denna Repaci
In the last few years, we have seen many new and powerful steganography and steganalysis techniques reported in the literature. In the following paper we go over some general concepts and ideas that apply to steganography and steganalysis. Specifically we establish a framework and define notion of security for a steganographic system. We show how conventional definitions do not really adequately cover image steganography and an provide alternate definition. We also review some of the more recent image steganography and steganalysis techniques.
I hold a PhD in Electrical and Computer Engineering from Binghamton University (NY) where I researched digital media security topics such as Steganography/Steganalysis or Watermarking using state-of-the-art Deep Learning and Machine Learning.
Deep learning has proven incredibly successful in a plethora of fields. In computer vision, deep neural networks are now the state-of-the-art for a variety of tasks. At the first glance, steganography and steganalysis appear to be very much different tasks than classical computer vision tasks, yet deep learning, especially convolutional neural networks, popularized by the computer vision field, have outperformed all classical feature-based approaches for detecting steganography. Intuitively, steganographic embedding changes are weak, noise-like signals executed primarily in complex content, such as textures and edges. Since computer vision classifies and categorizes content, it is also suitable for detecting the presence of noise-like stego signals modulated by content.In this dissertation, we focus on refactoring steganography detectors with more modern and general components, qualitatively understanding their strengths and failure cases, and using them algorithmically to improve steganography. First, we show that many custom ingredients long believed to be necessary for successfully training a deep neural network for steganography detection can be omitted in favor of more general-purpose convolutional architectures with very few domain-specific changes. Next, we focus on understanding what makes deep neural networks superior to their classical feature-based predecessors. Lastly, we use these powerful steganography detectors as a feedback loop in novel batch steganography algorithms, which allocate more payload in images where state-of-the-art detectors fail to detect steganography.
In this paper, we investigate pre-trained computervision deep architectures, such as the EfficientNet, MixNet, and ResNet for steganalysis. These models pre-trained on ImageNet can be rather quickly refined for JPEG steganalysis while offering significantly better performance than CNNs designed purposely for steganalysis, such as the SRNet, trained from scratch. We show how different architectures compare on the ALASKA II dataset. We demonstrate that avoiding pooling/stride in the first layers enables better performance, as noticed by other top competitors, which aligns with the design choices of many CNNs designed for steganalysis. We also show how pre-trained computer-vision deep architectures perform on the ALASKA I dataset.
Practical steganalysis inevitably involves the necessity to deal with a diverse cover source. In the JPEG domain, one key element of the diversification is the JPEG quality factor, or, more generally, the JPEG quantization table used for compression. This paper investigates experimentally the scalability of various steganalysis detectors w.r.t. JPEG quality. In particular, we report that CNN detectors as well as older feature-based detectors have the capacity to contain the complexity of multiple JPEG quality factors within a single model when the quality factors are properly grouped based on their quantization tables. Detectors trained on multiple JPEG qualities show no loss of detection accuracy when compared with dedicated detectors trained for a specific JPEG quality factor. We also demonstrate that CNNs (but not so much feature-based classifiers) trained on multiple qualities can generalize to unseen custom quantization tables compared to detectors trained for specific JPEG qualities. Their ability to generalize to very different quantization tables, however, remains a challenging task. A semi-metric comparing quantization tables is introduced and used to interpret our results.
All articles published by MDPI are made immediately available worldwide under an open access license. No specialpermission is required to reuse all or part of the article published by MDPI, including figures and tables. Forarticles published under an open access Creative Common CC BY license, any part of the article may be reused withoutpermission provided that the original article is clearly cited. For more information, please refer to
Feature papers represent the most advanced research with significant potential for high impact in the field. A FeaturePaper should be a substantial original Article that involves several techniques or approaches, provides an outlook forfuture research directions and describes possible research applications.
Abstract:Image steganography is applied to hide some secret information. Occasionally, steganography is used for malicious purposes to hide inappropriate information. In this paper, a new deep neural network was proposed to detect context-aware steganography techniques. In the proposed scheme, a high-boost filter was applied to alleviate the high-frequency while retaining the low-frequency details. The high-boost image was processed by thirty SRM high-pass filters to obtain thirty high-boost SRM filtered images. In the proposed CNN, two skip connections were used to collect information from multiple connections simultaneously. A clipped ReLU layer was considered in spite of the general ReLU layer. In constructing the CNN, a bottleneck approach was followed for an effective convolution. Only a single global average pooling layer was used to retain the complete flow of information. SVM was utilized instead of the softmax classifier to improve the detection accuracy. In the experimental results, the proposed technique was better than the existing techniques in terms of the detection accuracy and computational cost. The proposed scheme was verified on BOWS2 and BOSSBase datasets for the HILL, S-UNIWARD, and WOW context-aware steganography algorithms.Keywords: image steganography; image steganalysis; convolutional neural network; deep learning network
Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
The evolvement in digital media and information technology over the past decades have purveyed the internet to be an effectual medium for the exchange of data and communication. With the advent of technology, the data has become susceptible to mismanagement and exploitation. This led to the emergence of Internet Security frameworks like Information hiding and detection. Examples of domains of Information hiding and detection are Steganography and steganalysis respectively. This work focus on addressing possible security breaches using Internet security framework like Information hiding and techniques to identify the presence of a breach. The work involves the use of Blind steganalysis technique with the concept of Machine Learning incorporated into it. The work is done using the Joint Photographic Expert Group (JPEG) format because of its wide use for transmission over the Internet. Stego (embedded) images are created for evaluation by randomly embedding a text message into the image. The concept of calibration is used to retrieve an estimate of the cover (clean) image for analysis. The embedding is done with four different steganographic schemes in both spatial and transform domain namely LSB Matching and LSB Replacement, Pixel Value Differencing and F5. After the embedding of data with random percentages, the first order, the second order, the extended Discrete Cosine Transform (DCT) and Markov features are extracted for steganalysis.The above features are a combination of interblock and intra block dependencies. They had been considered in this paper to eliminate the drawback of each one of them, if considered separately. Dimensionality reduction is applied to the features using Principal Component Analysis (PCA). Block based technique had been used in the images for better accuracy of results. The technique of machine learning is added by using classifiers to differentiate the stego image from a cover image. A comparative study had been during with the classifier names Support Vector Machine and its evolutionary counterpart using Particle Swarm Optimization. The idea of cross validation had also been used in this work for better accuracy of results. Further parameters used in the process are the four different types of sampling namely linear, shuffled, stratified and automatic and the six different kernels used in classification specifically dot, multi-quadratic, epanechnikov, radial and ANOVA to identify what combination would yield a better result.
With the advent of technology and digitalization of services, several crucial data and personal data have been vulnerable to cyber-attacks. The bank transactions, communication between government organizations or personnel, national defense units etc. generate high amount of critical information which demand security systems to prevent loss or corruption of data. Steganography and Cryptography are two frequently used information security techniques for efficient and confidential communication. Several instances in real life use multiple methods for a resilient security system.
The confidential hiding of data for communication between a sender and a designated receiver is called steganography. This incognito data transfer can be done using copious types of multimedia formats like text, image, audio video and animation. The process of steganography consists of two algorithms-embedding and extraction. The embedding algorithm embeds or conceals the data into the medium that is used for the communication. In this paper, the medium used are JPEG images. JPEG is one of the most used formats as it is a standard image format used in scanners, photography, and other image processing tools.JPEG format is used in this paper due to ideal property of lossy compression which is necessary in critical information exchange. Hence, the images further to the embedding process will be called the stego image and it is transmitted using a transmission channel or the communication medium. Steganalysis which is a technique that intents to detect the presence of a hidden message is performed by a steganalyst at the communication channel. If the messages surpass the steganalyst, it reaches the receiver where the message is retrieved using the extraction algorithm. Like cryptography, the technique of steganography also uses a key which is shared only between the sender and receiver. The key is the most vital entity for both embedding and extraction algorithms1. The process is explained in Fig. 1. The image embedding in the paper is performed in two different domains-Spatial domain and Transform domain. The embedding in spatial domain is done right on the picture pixels whereas in the Transform domain the image is transformed using Discrete Cosine Transform to the coefficients. In the work, the steganographic algorithms used in the spatial domain are Least Significant Bit Matching (LSB M), Least Significant Bit Replacement (LSBR) and Pixel Value Differencing (PVD). F5 steganographic scheme is used in the transform domain.
7fc3f7cf58