MFeRNet: A Deep CNN Approach for Detecting Median Filter Tampering in Re-Compressed Images

The utilisation of median filtering (MF), a nonlinear signal processing technique, offers distinct advantages within picture anti-forensics. Consequently, there has been an increased focus on the forensic investigation of MF. However, due to lossy compression, identifying MF in the compressed domain is challenging. Towards this, research presents a novel approach for forensic analysis of MF in compressed images based on utilising deep noise residuals. In this framework, median filtering residuals (MFR) are employed to preprocess the images by passing through two streams. After that, the MFR output is extended to encompass two parallel blocks with different dilation rates to form a fusion feature vector. Further, the MFeRNet framework incorporates convolution, specifically developed to enhance information integration from several streams compared to conventional techniques. The proposed method, MFeRNet, aims to effectively integrate the three-level information of an image and comprehensively extract forensic clues in a compressed scenario. In addition, the experimental results demonstrate that the proposed methodology exhibits superior performance and reduced training time compared to the early reported techniques with equivalent convolution depth.