Image tampering detection for forensics applications

The rapid growth of image editing softwares has given rise to large amounts of doctored images circulating in our daily lives, generating a great demand for automatic forgery detection algorithms in order to determine the authenticity of a candidate image in a timely fashion. A good forgery detection algorithm should be passive and blind, requiring no extra prior knowledge of the image content or any embedded watermarks. By analyzing the abnormal behaviors of doctored images from authentic images, one can design forgery detectors based on a collection of cues in the image formation process. In this thesis, we first present a fully automatic consistency checking algorithm for detecting arbitrarily-shaped splicing areas in a digital image. We specifically study the Camera Response Function (CRF), a fundamental property in cameras mapping input irradiance to output image intensity. A test image is first automatically segmented into distinct areas. One CRF is estimated from each area using geometric invariants from Locally Planar Irradiance Points (LPIPs). To classify a boundary segment between two areas as authentic or spliced, CRF-based cross fitting and local image features are computed and fed to statistical classifiers. Such segment-level scores are further fused to infer the image-level authenticity decision. Tests on two benchmark data sets reach performance levels of 70% precision and 70% recall, showing promising potential for real-world applications. Moreover, we examine individual features and discover the key factor in splicing detection. Our experiments show that the anomaly introduced around splicing boundaries plays the major role in successful detection. Such finding is important for designing effective and efficient solutions to image splicing detection. As for the second focus of this thesis, we move beyond single forgery detector and propose a universal framework to integrate outputs from multiple detectors. Multiple cue fusion provides promises for improving the detection robustness, however has never been systematically studied before. By fusing multiple cues, the tampering detection process does not rely entirely on a single detector and hence can be robust in face of missing or unreliable detectors. We propose a statistical fusion framework based on Discriminative Random Fields (DRF) to integrate multiple cues suitable for forgery detection, such as double quantization artifacts and camera response function inconsistency. The detection results using individual cues are used as observations from which the DRF model parameters and the most likely node labels are inferred indicating whether a local block belongs to the tampered foreground or the authentic background. Such inference results also provide information about localization of the suspect spliced regions. The proposed framework is effective and general - outperforming individual detectors over systematic evaluation and easily extensible to other detectors using different cues. Both the consistency checking and multiple cue fusion frameworks are highly flexible, ready to accommodate other cues. The contribution of this thesis is therefore not limited to workable, powerful algorithms for forgery detection, but more importantly generalizable strategies in the design of potential forgery detection modules that might arise in the future.