Image Quality Metrics Based Blind Digital Image Forensics

With the development of photography, it is easier and cheaper for people to take a photo today, so photos gradually become a common way for people to record their daily life. Besides, the developed social media provide a convenient platform for digital photos to be shared. As a result, billions of new digital images are uploaded to the Internet every day, which may become sources of image forgery. In addition, the updated image editors simplify the image tampering tools and show excellent performance. All above are threatening the information security. Digital image forgery issues happened more frequently in recent years and resulted in detrimental effects, so it is meaningful to study on image forensics.In this paper, we focus on the most popular blind digital image forensics research and study on its two main research areas which are source camera identification and tampering detection. As the first step of forensics, source camera identification is to verify the reliability of digital images and to provide evidences for the following detection tasks as well. Previous approaches for source camera identification and tampering detection mainly study on the noise pattern and tampering means separately, which are not generalized. So, in this paper, we associate the image quality metrics with image forensics to generate more practical and universal forensics methods. The main work is that by analyzing the impact of image formation and image tampering separately on image quality, we propose several no-reference image quality features on forensics, and the model to identify source cameras and to locate tampered areas separately as well. The major contributions are as follows:(1) To solve the problem of identifying outlier models of the training sets, we propose a classifier combination based source identification strategy for digital camera images. In the proposed classification structure, a one-class classifier is orderly used in the framework together with a multi-class classifier to determine the outlier. Besides, we analyze the theoretical error brought by the cascade system, providing convenience for future work. At the same time, we analyze the correlation between image pixels with their neighborhoods according to the differences between qualities of images from various source cameras, resulting in corresponding quality assessment features which could be used for learning the classification models. In the experiment part, we show that the performance of the proposed method is better than previous works.(2) To generate a relatively universal way which can detect image tampering without many constraints, we utilize the image quality assessment theories to propose a no-reference quality metrics based image tampering detection method, including 13 features based on the statistics in both spatial domain and frequency domain of an image and the models to detect global tampering and locate local tampering separately using these features. We conduct simulations on common tampering means including compression and blur, and the accuracy results of global detection are all over 96% which is higher than other related methods. In the localization experiment we find that the detection maps of some images are not very accurate, so in some extra experiments we change the slide window size and eventually get satisfied performance. Specially, we tamper some images by Photoshop which is able to make fake images hardly to be recognized, and the detection results are closed to the ground truth, verifying the effectiveness of the proposed approach in the real and complex application scenarios.