Year 2022-2025

Abstract: When a sudden public event occurs, a large number of deliberately tampered images spread on social networks, seriously affecting social stability. At present, most of the image forensic algorithms conducted the statistical model analysis for a single image operation or adopted the end-to-end deep learning framework for general forensic analysis. These methods are still unable to cope with the actual network transmission scene. Toward the actual network communication environment, this project intends to start from the application of tampering detection of single suspicious public opinion image spreading on social networks and combine the evidence fusion theory with setting up an image forensic model. After uploading images, each social network platform often performs a series of operations such as size reduction and JPEG compression. These operations cannot be regarded as malicious tampering, and at the same time, it will cover up the original image statistical characteristics. This project focuses on the research of image forensics under the interference of the later operation of the social network platform. The main research contents include: 1) tracing the evidence of suspicious image social network, 2) estimation of the operation process of the social network platform, 3) the confidence assessment of image forensics algorithms based on local features and later operation process, 4) improvement of the statistical model of image forensics algorithms considering the influence of later operation, and 5) the set up of integrated image forensic model based on evidence fusion theory via combining the confidence level and the local feature constraint of each forensic tool. The development of this project is expected to enrich the existing image forensic model and strongly support the credibility of electronic evidence in forensic identification.

 

Year 2018 - 2020

Abstract: Image forensics is a frontier research topic in the field of information security. This project uses the high-order statistical-based noise estimation method as the entry point to study the new theories and methods of noise estimation and its forensic applications. The detailed works are as follows. We will analyze the mechanism of the scale-invariant kurtosis property, and investigate the influence of the differences in terms of the content and regular texture on stability of the kurtosis invariance. We will use the deep learning theory to extract the stable region with the best scale-invariant kurtosis property, and construct a multi-constrained cost function to determine the global variance of the image. We will conduct the box filter technique to extent the global estimation method to pixel-level local method. After constructing a weighted mechanism for the confidence level of samples for estimation based on the local noise estimation, we will explore the new method to estimate the noise level function under the condition of uneven distribution of samples. After deriving the relationship model between the noise level function, the camera response function and the external adjusted curve, we will estimate the external curve based on the Bayesian theory. We will combine the features of spectral residual and the guided filtering error to measure the homogeneity of each pixel, and explore the monotone relation between the homogeneity and the noise estimation error. We will further propose an optimized samples selection and modification approach based on a stepwise updating strategy, and then estimate the maximum likelihood noise distribution for an image and the probability of each pixel obeying the distribution using the theories of sparse representation and maximum likelihood. The results of the project are of great significance to the theories and techniques of digital forensics and multimedia information security.