General-purpose image forensics using patch likelihood under image statistical models

This paper proposes a new, conceptually simple and effective forensic method to address both the generality and the fine-grained tampering localization problems of image forensics. Corresponding to each kind of image operation, a rich GMM (Gaussian Mixture Model) is learned as the image statistical model for small image patches. Thereafter, the binary classification problem, whether a given image block has been previously processed, can be solved by comparing the average patch log-likelihood values calculated on overlapping image patches under different GMMs of original and processed images. With comparisons to a powerful steganalytic feature, experimental results demonstrate the efficiency of the proposed method, for multiple image operations, on whole images and small blocks.