Selection of JPEG steganography algorithms using a feature based model

Abstract

JPEG image steganographic techniques use the DCT coefficients scaled by quantization table to make secure data hiding without degrading the image quality. The selection process of data embedding locations in lower frequency DCT coefficients should be carefully considered in each image blocks as these lower frequency coefficients are high sensitive to human eyes. Some of the existing related JPEG steganographic methods have been proposed with primary quantization table modification to hide message bits in the quantized DCT coefficients with minimal distortion by analyzing the properties of quantization table entry and relevant DCT coefficients. The performance of the JPEG steganographic methods is evaluated by the imperceptibility and embedding capacity. In the literature of quantization table modification based JPEG steganography, the middle frequency coefficients in each image block are utilized to embed maximum message size by modifying the middle part of the relevant quantization table values with minimizing the effect of visual perception. However, the data hiding techniques in lower frequency coefficients from the existing studies endure from imperceptibility while increasing the message size. This study suggests the lower frequency data hiding algorithms with utilizing middle frequency data hiding in terms of the modification of lower and middle part of the quantization table values by evaluating image quality parameters and it doesn’t affect the perceptual detectability and improves embedding capacity. The proposed JPEG steganography investigates the modification of quantization table values with regarding to selected lower frequency DCT coefficients for data hiding and selects different data hiding patterns in lower frequency area in terms of modification of quantization table. Finally, it returns the pair of relevant modified quantization table and generated data hiding pattern for an image based on the empirical results of the PSNR values. The pair that contains modified quantization table and data hiding pattern shared by the sender is used as a secrete key to extract the message at the receiver side. From the preliminary studies, the selection of appropriate lower frequency coefficients in image block to hide the optimum size of secrete message with perceptual un-detectability is dependent on the combination of image features, message size and the hiding algorithm. Further, this study recommends a dynamic model to keep the consistency of the combination of image features, message size and the hiding algorithm in terms of quantization table modification and this model based steganography suggests a dynamic model to cover image statistics. Eventually, the model prevents visually perceptible changes for maximum embedding message bits. The proposed method achieves a good imperceptibility level and it is evaluated by the PSNR value range 30dB to 45dB and maximum message size more than 52 bits per block for the selected JPEG image dataset. The dynamic model fitted between the quantization tables and cover image statistics shows the statistical significance with the p-value 0.0007634 and the model generated between the data hiding pattern and statistical features of DCT coefficients shows the statistical significance with the p-value 4.598e-13. The dynamic model for the selected data hiding patterns in the lower frequency coefficients hides the message and it is stego invariant for message analyzers.