Research On Digital Image Manipulation Forensics And Its Security

Digital multimedia such as images and videos are important information sources in people’s daily life.However,the development of image processing techniques and the popularity of the deep learning approach have led to the maturity of image-editing software and the increasing fidelity of forged images.This fact makes people strongly question the originality and authenticity of images and brings a severe test to image forensics.In particular,tamperers often use embellishment operations such as filtering,enhancement,and compression to erase traces left by tampering or enhance the visual quality of images.Therefore,the research on image manipulation forensics is of great significance.Considering the weaknesses of existing forensic methods,such as poor performance on low-quality and small-sized images,and lack of theoretical basis,from the perspective of capturing manipulation traces,combined with theoretical analysis and feature mining techniques,we propose to conduct an in-depth study of the forensics of several different image manipulations and improve the performance.Moreover,we also propose to test and improve the security of image manipulation forensics.The main contributions are as follows.(1)We propose a robust median filtering detection method based on the difference of frequency residuals.To solve the problem that existing methods are less robust for JPEG compression,we propose to extract the feature based on the difference of frequency residuals from JPEG images,and then distinguish between JPEG compressed unfiltered and median filtered images.Specifically,given that the frequency residual corresponding to the first filtering of the image is much larger than that of the second filtering,by performing two successive filtering operations on the test image,the feature of the difference in frequency residuals is extracted for median filtering detection.Moreover,an adaptive perturbation strategy is proposed to weaken the effect of image content on the feature,so as to determine a generic classification threshold.The proposed method uses a single threshold for the classification based on a single-dimensional feature,which is simple and efficient and has significant advantages over state-of-the-art methods in terms of the detection performance for low-quality and small-sized image blocks.(2)We propose a quantization step estimation algorithm for JPEG image forensics.Toaddress the problems that the existing estimation algorithms often lack theoretical foundation and have poor performance on low-quality and small-sized images,we propose to establish a response model between the probability density function of DCT coefficients of the decompressed image and the function of the candidate step.The candidate step leading to the largest response will be determined as the estimation result.The correspondence between the maximum response and the true quantization step is demonstrated,and thus a solid theoretical foundation is established for the estimation algorithm.Two fine adjustment measures based on some prior knowledge are further proposed to improve the estimation accuracy for lowquality and small-sized images.Experimental results show that the proposed algorithm outperforms some state-of-the-art methods,especially on low-quality and small-sized images.(3)We propose an improved error-based approach for detecting double JPEG compression with the same quantization matrix.To address the problems that the existing algorithms lack error distribution analysis and are weak in theory,we propose a new classification pattern for JPEG error blocks to locate almost identical truncation error blocks in single-and double-compressed images,which can accurately capture the effective error differences in images.Then,the distributions of rounding errors and truncation errors are analyzed to dig into the intrinsic causes of error differences at the theoretical level,and to extract highly discriminative statistical features for detecting double JPEG compressed images.Experimental results show that compared to some advanced methods,the proposed approach achieves better performance,especially on low-quality images.(4)Considering the security of digital image manipulation forensics,focusing on deep learning-based forensics models,we propose an anti-forensics strategy based on increased-confidence adversarial examples and a secure detection model based on random feature selection and gradient orthogonalization.On the one hand,in order to explore the security of deep forensics models,we propose to increase the attack strength and hence enhance the transferability of the attack by adjusting the confidence of adversarial examples.On the other hand,since the deep forensics model is highly vulnerable to adversarial examples,to improve its security,a new loss function is defined to maximize the angle between the gradients of the image over the model with full features and the model with random features.A threat model is further established to measure the model security in various aspects.Ex-perimental results show that,for image manipulation forensics tasks,the proposed secure forensics model achieves better defense performance against different kinds of adversarial examples under different attack scenarios.