A Median Filtering Forensics CNN Approach Based on Local Binary Pattern

Median filtering forensics has gradually become a research hotspot because of the wide application of the traditional median filtering (MF) in image tampering and anti-forensics. The difficulty of traditional median filtering based on machine learning forensics is feature extraction which is a manual selection process, and the Convolutional Neural Network (CNN) cannot also well perceive the traces left by median filtering straightly. By taking the Local Binary Pattern (LBP) data of an image as a presentation of the streaking artifact that is a very strong indication for median filtering, a median filtering forensics CNN approach with LBP is proposed, which can automatically learn and get median filtering features from image. Different from traditional CNN, an LBP perception layer is added before the following deep learning layers. Then, five convolution layers constitute the feature extraction group. Finally, the classifier group is composed of three full connected layers to decide whether the image is median filtered or not. The proposed approach is tested on several experiments and the experimental results demonstrate its effectiveness.