Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
This work addresses the task of detecting the tampering of images in which multiple images are combined into one (image splicing). To this end the authors introduce a novel generative adversarial training technique consiting of four models that are trained simultaneously. The method is shown to perform very well on existing data sets and a novel large scale image manipulation dataset. The experiments are detailed and code and data set will be made available. Three expert reviewers initially assessed the work as 8/7/5, and the authors provided a detailed rebuttal; two reviewers acknowledge reading the rebuttal and one adjusted the score, for a final rating of 8/7/6. Remaining concerns are around clarity of writing and limited analysis studying the high effectiveness of the method in more detail. Overall, this is an important problem and the paper makes a clear contribution, both in terms of an interesting novel method and a new data set.