Steganalysis for palette-based images using generalized difference image and color correlogram

In this paper, we attempt to propose a novel blind steganalysis algorithm for palette-based images. First, the generalized difference images between adjacent pixels were constructed, and then the moments of characteristic functions of difference images' histograms were extracted as features. Second, in order to measure the dependencies of neighboring colors, color correlogram technique is used to capture the global distribution of local spatial correlation of colors. The center of mass of the characteristic function of color correlogram and the absolute moments of autocorrelogram were extracted. Total of 13 dimension features were classified with machine learning technique. Number of experiments on several existing GIF steganography algorithms indicated that the proposed scheme is effective and gets good performance, especially when the embedding rate is not less than 20%. Experimental results also show that the average accuracy of our proposed scheme for different GIF steganography algorithms outperforms Lyu's algorithm more than 20%. It also showed that the proposed scheme achieved similar performance with Fridrich's scheme and higher accuracies comparing to Du's algorithm and biologically inspired features.

► The effect of data hiding on neighboring colors in palette-based images was studied.
► Image difference and color correlogram were used to capture the local correlation.
► High-order statistic was extracted to distinguish the cover and stego images.
► Better performance was obtained from numbers of experiments on palette-based images.