Joint-channel modeling to attack QIM steganography

Abstract

This paper presents a statistical steganalysis framework to attack quantization index modulation (QIM) based steganography. The quantization process is generally modeled as an additive noise channel. The proposed method exploits the fact that plain-quantization (quantization without message embedding) decreases local-randomness (or increases local-correlation) in the resulting quantized image. Moreover, QIM-stego image exhibits relatively higher level of local-randomness than the corresponding quantized-cover, though both are obtained using same set of parameters. The local-randomness (or inter-block correlation) of the test-image is used to capture traces left behind by quantization. A parametric model is developed to characterize channel-dependent local-randomness. Maximum likelihood estimation (MLE) framework is used to estimate parameters of the distribution of local-randomness mask. Distributions of parameters, estimated from the quantized-cover and the QIM-stego images, are used to characterize quantization with and without message embedding. To investigate variations in the estimated parameters as a function of frequency, inter- (variations within each channel(or subband)) and intra-channel (across all channels), joint-channel modeling and single-channel modeling, respectively, is considered. For each approach, a set of parametric detectors based on generalized likelihood ratio test (GLRT) is used to distinguish between the cover- and the stego-images. To improve detection accuracy, decisions from both detectors are fused to generate the final stego-detection decision. Effectiveness of the proposed framework is evaluated using a dataset consisting of over 35000 test-images obtained from 880 uncompressed natural images. Experimental results show that the proposed scheme can successfully detect QIM-stego images with very low false rates. In addition, performance comparison with existing state-of-the-art also shows that the proposed method performs significantly superior than the selected methods.