Median filtering forensics using multiple models in residual domain

Abstract

Median filtering, due to highly non-linear and content-preserving, has widely used in the multimedia security fields, such as anti-forensics, steganography, and steganalysis. In the past decade, many excellent algorithms have been proposed, and have solved median filtering forensics perfectly for images without post-processing. However, it is still a challenging task when the test images are of low resolution and post-processed by lossy JPEG compression. Moreover, in order to meet the requirements of detecting forgery from the massive image database in the social network, a fast and efficient detector is needed. In this paper, we present an improved method for median filtering forensics in challenging problems, such as detecting median filtering from a heavily JPEG compressed image. To do this, we first investigate median filtering traces by a group of diverse residuals. These residuals not only eliminate influence from image contents but also highlight the traces from different aspects. We then construct a feature set on the residuals by incorporating the Markov chain model with the auto-regressive model. These two compensated models effectively reveal different relationships among neighboring residuals. For a fast detector, a series of dimension reduction methods are employed. The experimental results on a composite image database demonstrate that the proposed method outperforms prior arts when detecting median filtering in heavily JPEG compressed images as well as low-resolution images.