Detecting median filtering via two-dimensional AR models of multiple filtered residuals

Median filtering, being an order statistic filtering, has been widely used in image denoising and recently also in image anti-forensics and anti-steganalysis. In the past few years, several methods have been developed for median filtering detection. However, it is still a challenging task to detect median filtering in JPEG compressed images. In this paper, we propose a novel method to solve this challenging task. We first generate median filtered residual (MFR), average filtered residual (AFR) and Gaussian filtered residual (GFR) by calculating the differences between an original image and its filtered images. Then, we propose to use two-dimensional autoregressive (2D-AR) model to characterize MFR, AFR and GFR separately, and further combine the 2D-AR coefficients of these three residuals into a set of features. Finally, the extracted feature set is fed into a support vector machine classifier for training and detection. Extensive experiments have demonstrated that compared with existing methods, the proposed one can achieve a considerable improvement in detecting median filtering in heavily compressed images.