Deep Residual Learning Using Data Augmentation for Median Filtering Forensics of Digital Images

Abstract

This paper addresses the median filtering forensics for a lossy compressed image with low resolution, which is essential for the identification of fake images and fake videos. A deep residual model with training data augmentation is employed in the proposed method. To solve the dilemma that the low-resolution image is the lack of enough statistical pixels for extracting reliable features, we propose a filter layer to widen the inputs for the convolutional neural network (CNN). First, we perform the high-pass filtering to an image in the filtered layer and stack the multiple filtered residuals into 16-channel feature maps as inputs of CNN. Then, a deep residual CNN model has proposed to self-learn the median filtering traces that are hidden in the JPEG lossy compressed image. To alleviate the over-fitting issue of the deeper CNN model, we employ a data augmentation scheme in the training to increase the diversity of training data and, thus, obtain a more stable median filtering detector. The experimental results demonstrate that the proposed net with training data augmentation outperforms state of the arts in both baseline test and generalization ability test, achieving at least 2% higher in terms of detection accuracy.