Data-Dependent Scaling of CNN’s First Layer for Improved Image Manipulation Detection

Abstract

Convolutional Neural Networks (CNNs) have become an effective tool to detect image manipulation operations, e.g., noise addition, median filtering and JPEG compression. In this paper, we propose a simple and practical method for adjusting the CNN's first layer, based on a proper scaling of first-layer filters with a data-dependent approach. The key idea is to keep the stability of the variance of data flow in a CNN. We also present studies on the output variance for convolutional filter, which are the basis of our proposed scaling. The proposed method can cope well with different first-layer initialization algorithms and different CNN architectures. The experiments are performed with two challenging forensic problems, i.e., a multi-class classification problem of a group of manipulation operations and a binary detection problem of JPEG compression with high quality factor, both on relatively small image patches. Experimental results show the utility of our method with a noticeable and consistent performance improvement after scaling.