An Improved Method of Median Filtering Forensics for Enhanced Image Security Detection

With the development of image processing technology, the forgers usually use median filtering to make their fakes appear more realistic because median filtering is a non-linear digital filtering technique which can preserve edges and smooth regions within an image. Therefore, the use of median forensics technology to identify the authenticity of images and ensure the security of image data has attracted everyone’s attention. However, how to effectively detect the median filter of high JPEG compressed small-size images is still a challenge in median filter forensics. In this paper, we proposed a framework based on deep residual learning to address this challenge. Specifically, a new convolutional neural network called MFFNet was constructed. In the first step, in order to extract the features left by the median filter, we innovatively designed a preprocessing layer containing various residuals to capture different median filter artifacts. Then, we elaborately designed a MFFNet to self-learn rich hierarchical features left in the highly JPEG compressed image for further classification. In order to prevent the over-fitting problem of the deep network, we adopted a series of enhancement schemes in the training stage to enrich the diversity of training data and obtain a more generateable and stable median filter detector. A large number of experimental results on the composite database show that the proposed framework significantly improves the detection performance compared to the latest methods for detecting highly small size image with JPEG compression.