Cyber Trust - Restore the Trustworthiness of Digital Photographs: Blind Detection of Digital Photograph Tampering

值得信赖的照片在许多应用中发挥着重要作用，例如新闻报道，情报信息收集，刑事调查，安全监视以及医疗保健。然而，随着数字时代的到来，照片的可信度不再是理所当然的。本项目将开发一个完全盲的、被动的检测数码照片篡改的系统。不需要额外的加密、签名提取或信息嵌入过程。通过分析图像中的自然信号/场景特征，在检查点检测内容篡改操作。我们采用一种创新的方法，将信号处理和计算机图形技术相结合。该信号处理方法包括有效地利用高阶信号统计、信号采集设备建模、图像分解和图像结构分析来识别信号级的篡改伪影。计算机图形学方法包括三维几何估计、照明场恢复和场景重建等新技术，以检测场景级别的不一致性，如阴影、阴影和几何。我们的目标是一个成功的系统，使任何攻击机动尽可能困难。该系统还将同等地强调可靠性和信息性——将用地点和理由解释怀疑。为了实现更广泛的影响，建议的研究将包括部署公共图像伪造检测引擎，发布大型原始数据集，以及定义评估基准。

Trustworthy photographs play an important role in many applications such as news reporting, intelligence information gathering, criminal investigation, security surveillance, as well as health care. However, with the advent of digital age, the trustworthiness of pictures could no longer be taken for granted. This project will develop a completely blind and passive system for detecting digital photograph tampering. No extra encryption, signature extraction, or information embedding processes are needed. Content tampering operations are detected at the point of checking by analyzing the natural signal/scene characteristics in the image. We take an innovative approach integrating techniques from signal processing and computer graphics. The signal processing method involves effective use of higher-order signal statistics, signal acquisition device modeling, image decomposition, and image structural analysis to identify tampering artifacts at the signal level. The computer graphics approach includes novel techniques of 3D geometry estimation, illumination field recovery, and scene reconstruction to detect inconsistency at the scene level like shadows, shading, and geometry. We aim at a successful system that makes any attacking maneuver as difficult as possible. The system will also provide equal emphasis on robustness and informativeness - suspicions will be explained with locations and reasons. To achieve broader impacts, the proposed research will include deployment of a public image forgery detection engine, release of a large original data set, and definitions of evaluation benchmark.