Training CNNs in Presence of JPEG Compression: Multimedia Forensics vs Computer Vision

Convolutional Neural Networks (CNNs) have proved very accurate in multiple\ncomputer vision image classification tasks that required visual inspection in\nthe past (e.g., object recognition, face detection, etc.). Motivated by these\nastonishing results, researchers have also started using CNNs to cope with\nimage forensic problems (e.g., camera model identification, tampering\ndetection, etc.). However, in computer vision, image classification methods\ntypically rely on visual cues easily detectable by human eyes. Conversely,\nforensic solutions rely on almost invisible traces that are often very subtle\nand lie in the fine details of the image under analysis. For this reason,\ntraining a CNN to solve a forensic task requires some special care, as common\nprocessing operations (e.g., resampling, compression, etc.) can strongly hinder\nforensic traces. In this work, we focus on the effect that JPEG has on CNN\ntraining considering different computer vision and forensic image\nclassification problems. Specifically, we consider the issues that rise from\nJPEG compression and misalignment of the JPEG grid. We show that it is\nnecessary to consider these effects when generating a training dataset in order\nto properly train a forensic detector not losing generalization capability,\nwhereas it is almost possible to ignore these effects for computer vision\ntasks.\n