Information-theoretic bounds of digital image forensics

数字图像和图像处理渗透到日常生活中。这就需要新的技术来验证数字图像数据的真实性和完整性。这两个问题都与年轻的数字图像取证研究领域有关。已知的方法可以通过统计分析所涉及的图像处理算子的痕迹来检测图像伪造。然而，大多数已知的方法是启发式的，其有效性仅在实验室条件下为人所知。很少有严格的证据告诉我们一种方法为什么以及在什么条件下能产生可靠的结果。目前的发展越来越多地采用一种方法来结合更多的功能，并求助于机器学习来管理由此产生的复杂性。然而，这种方法对因果关系和限制几乎没有什么基本的见解。拟议的项目打破了这种方法。它试图为可用于图像取证的信息建立上限。重点放在已知方法不可靠的情况下。这一方法第一次告诉我们，进一步改进法医方法--启发式方法和理论上建立的方法--是否有希望；或者是否不存在更多对法医有用的痕迹。

Digital images and image processing pervade everyday life. This calls for new techniques to verify the authenticity and integrity of digital image data. Both questions pertain to the young research field of digital image forensics. Known methods can detect image forgeries by statistically analyzing traces of the involved image processing operators. However, most known methods are heuristic and their effectiveness is known for laboratory conditions only. Rigorous proofs telling us why and under what conditions a method produces reliable results are rare. Current developments increasingly follow an approach to combine ever more features and resort to machine learning for managing the resulting complexity. This approach, however, promises few fundamental insights into causal relationships and limitations. The proposed project breaks with this approach. It tries to establish upper bounds for the information that can be exploited for image forensics. Emphasis is put on situations where known methods are unreliable. This approach informs us for the first time on whether further refinements of forensic methods -- heuristic and theoretically founded ones alike -- are promising; or whether no more forensically useful traces exist.