Calibration-based features for JPEG steganalysis using multi-level filter

This paper proposes a novel feature set for JPEG steganalysis calibrated by multi-level filters. These features are built by the entries of co-occurrence matrices from multiple new planes obtained by subsampling the undecimated DCT coefficients according to specific element positions. The method can also be interpreted as a multi-level calibration method in the spatial domain by shifting the cropping positions in the original image to enhance the distinguishing ability of the feature set. Experimental results show that the proposed method can give the less detection error than that of some state-of-the-art steganalysis algorithms such as DCTR and CC-JRM.