Exposing image splicing with inconsistent sensor noise levels

Abstract

Splicing is a commonly used image tampering operation, where a part of one image is pasted into another image. The forged image can have completely different semantic from the original one and may mislead people in some serious occasions. To rebuild the credibility of the images, extensive forensic methods aiming to locate the spliced areas have been proposed in recent years. Among these methods, the noise based ones, which utilize the fact that images from different sources tend to have various noise levels, have drawn much attention due to their convenience to implement. However, most of the existing noise based methods are under the assumption that a synthetic additional white Gaussian noise (AWGN) is involved during the splicing. This maybe not the case in practice. In this study, we utilize the difference of the intrinsic sensor noise of the source images to expose the potential image splicing. In practice, the sensor noise level difference is common between images captured with different ISO settings. Through analyzing the characteristics of the sensor noise, a weighted noise level is proposed for reducing influences from image content thus can better localizing the splicing region. Specifically, the noise level of a questioned image is first estimated locally with principal component analysis (PCA)-based algorithm. Then, the estimated noise levels are weighted before clustering with k-means. The experimental results demonstrate the superiority of the proposed method over several state-of-the-art methods, not only for splicing localization purpose, but also for splicing detection purpose.