High-volume data hiding in images: Introducing perceptual criteria into quantization based embedding

Abstract

Information-theoretic analyses for data hiding prescribe em bedding the hidden data in the choice of quantizer for the host data. In this paper, we consider a suboptimal implementation of this prescription, with a view to hiding high volumes of data in images with low perceptual degradation. Our two main findings are as follows: (a) In order to limit perceptual distortion while hiding large amounts of data, the hiding scheme must use perceptual criteria in addition to information-theoretic guidelines. (b) By focusing on "benign" JPEG compression attacks, we are able to attain very high volumes of embedded data, comparable to information-theoretic capacity estimates for the more malicious Additive White Gaussian Noise (AWGN) attack channel, using relatively simple embedding techniques.

Steganographic capacity estimation for the statistical restoration framework

“…For DCT and discrete wavelet transform (DWT) domains, it is assumed that a distortion of ±1 for the coefficients in the chosen frequency band ensures perceptual transparency of the stego signal -with a proper choice of frequency domain terms for hiding, perceptual transparency for the DCT domain is shown in our prior work. 15,16 Instead of specifying a fixed value for the distortion constraint D 1 , our aim is to minimize the perceptual distortion during hiding while ensuring that the other constraints are maintained.…”

Steganalysis of quantization index modulation data hiding

“…Unlike standard quantization, these regions overlap for adjacent values of s. We note at this point that if the goal of the steganographer is to mimic an existing quantizer, for example a compression scheme, then the hider can stop here, without using dither modulation. In [4] and [7], the authors use this to imitate the output of JPEG and JPEG2000 respectively. We examine the detection of this first case below.…”

“…For example both [4] and [7] exclude low-valued coefficients from embedding, to avoid visual distortion of the final image. Furthermore, as we mentioned above the host data typically has a characteristically sharp increase of probability near the mean, which will be noticeably smoothed by hiding.…”

“…We assume that our classifier is targeted to a particular implementation of QIM, and choose the features to match. For our testing, we tested on a QIM implementation, [4], that embeds in the 8x8 blockwise DCT coefficents of an image. Since the DCT tends to decorrelate the intensity values, our model of an i.i.d.…”

“…Over the past few years, QIM-based data hiding has received increasing attention from the data hiding community because it is more robust than established techniques such as spread spectrum and least significant bit (LSB) hiding. Recently proposed QIM schemes include Chen and Wornell's QIM and dither modulation [2], Eggers et al's scalar Costa scheme (SCS) [3], and application tailored implementations such as [4,5,6,7].…”