Kwon Myung-Joon ; Nam Seung-Hun ; Yu In-Jae ; Lee Heung-Kyu ; Kim Changick

Abstract

Detecting and localizing image manipulation are necessary to countermalicious use of image editing techniques. Accordingly, it is essential todistinguish between authentic and tampered regions by analyzing intrinsicstatistics in an image. We focus on JPEG compression artifacts left duringimage acquisition and editing. We propose a convolutional neural network (CNN)that uses discrete cosine transform (DCT) coefficients, where compressionartifacts remain, to localize image manipulation. Standard CNNs cannot learnthe distribution of DCT coefficients because the convolution throws away thespatial coordinates, which are essential for DCT coefficients. We illustratehow to design and train a neural network that can learn the distribution of DCTcoefficients. Furthermore, we introduce Compression Artifact Tracing Network(CAT-Net) that jointly uses image acquisition artifacts and compressionartifacts. It significantly outperforms traditional and deep neuralnetwork-based methods in detecting and localizing tampered regions.