Robust median filtering forensics based on the autoregressive model of median filtered residual

One important aspect of multimedia forensics is exposing an image's processing history. Median filtering is a popular noise removal and image enhancement tool. It is also an effective tool in anti-forensics recently. An image is usually saved in a compressed format such as the JPEG format. The forensic detection of median filtering from a JPEG compressed image remains challenging, because typical filter characteristics are suppressed by JPEG quantization and blocking artifacts. In this paper, we introduce a robust median filtering detection scheme based on the autoregressive model of median filtered residual. Median filtering is first applied on a test image and the difference between the initial image and the filtered output image is called the median filtered residual (MFR). The MFR is used as the forensic fingerprint. Thus, the interference from the image edge and texture, which is regarded as a limitation of the existing forensic methods, can be reduced. Because the overlapped window filtering introduces correlation among the pixels of MFR, an autoregressive (AR) model of the MFR is calculated and the AR coefficients are used by a support vector machine (SVM) for classification. Experimental results show that the proposed median filtering detection method is very robust to JPEG post-compression with a quality factor as low as 30. It distinguishes well between median filtering and other manipulations, such as Gaussian filtering, average filtering, and rescaling and performs well on low-resolution images of size 32 × 32. The proposed method achieves not only much better performance than the existing state-of-the-art methods, but also has very small dimension of feature, i.e., 10-D.