Countering Universal Image Tampering Detection with Histogram Restoration

In this paper, we point out state-of-the-art algorithm in natural image splicing detection, namely the transition probability matrix feature proposed by Shi, et al., can be attacked by modifying block discrete cosine transform (BDCT) coefficients without significantly degrading quality of the spliced image. BDCT coefficients of the spliced image are modified so that its distance to a close authentic image in feature space is minimized. The minimization is accomplished with a greedy algorithm. The modification makes the spliced image statistically similar to the authentic image so as to reduce the effectiveness of detection algorithm. The performance of the algorithm is evaluated on Columbia Image Splicing Detection Evaluation Dataset. With the proposed anti-forensics post processing, detection accuracy and true positive rate reduces to 69.4% and 62.5% respectively, while the processed images still maintain average peak signal-to-noise ratio (PSNR) at 42.22db.