Development of cyber–crime detection software for Re–compression based Multimedia Forensics

With the availability of immoderate and powerful editing software, re–compression based artifacts gaining high popularity in image manipulation where the quality factor is different between forged and unforged regions. This difference in compression factor can be exploited to detect and localize the tampering. We present a forensic technique to detect inconsistencies in JPEG compression ratios over different regions of an image. Re–compression based forgeries can be detected by analyzing the proper artifacts introduced while manipulating the image. Classification of an image into forged and unforged class require training with appropriate features. As the feature selection and extraction contribute high complexity. We achieve this is performing a JPEG block–wise Convolution Neural Network (CNN) classification, applied to our test images. CNN is a biologically inspired feed forward neural network which segments the whole image into patches/ blocks and efficiently learn those patches. These overlapping blocks are considered one–by–one using movement of stride. The performance of the proposed forgery localization method has been maximized, by considering vertical and horizontal strides of magnitude, as low as eight pixels. This helped us achieve extremely minute forgery localization units, hence increasing the detection accuracy highly. To automate the detection of re–compression based forgery in JPEG, we aim to develop a software tool in which the image analysis take place. The software consist of two modules, pre–processing of an image and detection of tampered regions using Convolution Neural Network (CNN) model. In the first module, the images are divided into overlapping blocks followed by features extraction. In the second module, these features fed into CNN for learning and classification of an image into forged and unforged images