Image provenance inference through content-based device fingerprint analysis

We have introduced different intrinsic device fingerprints and their applications in image provenance inference. Although with varying levels of accuracy, the device fingerprints arising from optical aberration, CFA interpolation, CRF, and in-device image compression are effective in differentiating devices of different brands or models. Although they cannot uniquely identify the source device of an image, they do provide useful information about the image provenance and are effective at narrowing down the image source to a smaller set of possible devices. More than half of the chapter was spent on SPN, which is the only fingerprint that distinguishes devices of the same model. Because of its merits, such as the uniqueness to individual device and the robustness against common image operations, it has attracted much attention from researches and been successfully used for source device identification, device linking, source-oriented image clustering, and image forgery detection. In spite of the effectiveness of SPN, it is by nature a very weak signal and may have been contaminated by image content and other interferences. Its successful application requires jointly processing a large number of pixels, which results in very high dimensionality of SPN. This may bring huge difficulties in practice, e.g., in large-scale source-oriented image clustering based on SPN, so it is essential to conduct research on the compact representation of SPN for fast search and clustering.