Research On Image Splicing Forensics Based On Markov Feature And Srm Steganalysis Model

As a kind of electronic evidence,digital images play an important role in proving the truth.With the development of computer technology and artificial intelligence technology,the authenticity of the image has been seriously challenged,and the authenticity inspection of the image has become a new research field.There are many ways to tamper images,and their fundamental purpose is to change the original content of the image and gain some benefits.Image splicing is one of the image manipulation methods.The operation is simple and fast.After the post-processing,people can hardly judge the authenticity.In recent years,domestic and foreign scholars have proposed a series of methods for detecting image splicing operations and achieved good results,However,the robustness of the splicing detection algorithm and the accuracy of the detection of the splicing region have yet to be improved.The natural statistical model based on Markov features is currently one of the best methods for image splicing detection.After summarizing the innovative algorithms proposed by previous researchers,Improvements have been made by feature extraction methods,Markov feature calculations,and image preprocessing.Multi-residual Markov splicing detection model based on SRM model,Image-Splicing Detection by color DCT quantification Markov model,and image splicing region location model based on convolutional neural network have been proposed.The details are as follows:1? Multi-residual Markov splicing detection model based on SRM modelAiming at the problem of single calculating the difference matrix for the traditional Markov feature is not robust to the splicing detection.This paper analyzes the feasibility of Steganalysis Rich Model(SRM)applied to image splicing detection.It is also found in the experiment that converting a color image to a grayscale image and then feature extraction will possibly causing a loss of information.Based on those two points,a multi-residual types in SRM are introduced to improve the traditional Markov characteristics.Which respectively extracts 10 different types of Markov features from three color channels,and 30 unique SVM classifier are trained to make the classification through proposed decision-making algorithm.This method achieves the accuracy of 95.40% at Columbia image splicing detection evaluation dataset and is robust to post-processing operations such as JPEG compression and median filtering.2? Image-Splicing Detection by Color DCT Quantification Markov ModelDue to the high time complexity when multiple transition probability matrices are calculated for multi-residual Markov features.we presents a color DCT quantification Markov model for splicing detection inspired by JPEG compression algorithm.Each image is quantified in DCT domain in order to highlight the high frequency information,and then Markov features would be extracted from three color channels respectively.Finally,three SVM classifier are trained to make the classification through proposed decision-making algorithm.This method achieves a high accuracy of 99.94% in IFS-TC training set.3? Image-Splicing Region Location Model based on Convolutional Neural NetworkTraditional splicing detection algorithms need to construct features manually.This phase cost researchers a lot of time and effort.Moreover,with the development of image processing technology and artificial intelligence technology,image tamper technology is rising too,and the artificially constructed features are not robust enough.It is also getting harder to locate the tampered region.This paper proposes constructing a convolutional neural network using a fixed pre-convolution kernel for detecting and locating the splicing regions by feature self-learning.In the research,it was found that convolutional neural network(CNN)models can be used for feature self-learning in tampered regions.By adding high-pass filtered convolutional kernels in front of the network and using exponential linear units as activation functions,convolutional neural networks can be trained to learn the features from image splicing tampered region.This method has a high accuracy of 99.94% in IFS-TC training set.The first proposed model is basically the same as the second model.All of them improve the accuracy of image splicing detection by improving the traditional Markov characteristics.The former has a high time complexity,the latter is faster due to the reduced number of features.Both models can only detect the splicing image and cannot locate the specific tampering region.Image-Splicing Region Location Model based on Convolutional Neural Network compensates for the deficiency of the first two models and achieves the positioning of the tampered region.