New Distortion Metric for Efficient JPEG Steganography Using Linear Prediction

Abstract

Hiding a secret message in a cover image with the least possible statistical detectability is the objective of steganography. This is generally formulated as a problem of minimal-distortion embedding and practically implemented by incorporating an efficient coding method and a well-designed distortion metric. In this paper, we construct a new distortion metric for JPEG steganography, which employs a local linear predictor to generate both the intra- and inter-block prediction errors of a quantized DCT coefficeint, and then accumulates them to form the distortion metric for this coefficient. Such distortion metric is then further integrated in the minimal-distortion framework using STC to give rise to the proposed JPEG steganographic scheme. This scheme exploits the proposed distortion metric to guide the STC to hide the secret message in those quantized DCT coefficients with minimal distortion cost. Consequently, the average changes of both first- and second-order statistics of quantized DCT coefficients and thus the statistical detectability would decrease significantly. Compared with prior arts, experimental results demonstrate the effectiveness of the proposed scheme in terms of secure embedding capacity against steganalysis.