Identifying Image Origin Via Low Dimensional PRNU Features

Everyone can easily edit and tamper with images by powerfull digital image processing software.Therefore,image source identification of digital image forensics has become an urgent problem.Most current image source identification algorithms are based on sensor pattern noise features as an inherent fingerprint/feature of image source identification.However,current camera fingerprints,used for identification,with the high dimensional features indeed become the obstacle of designing an efficient and reliable classifier.Therefore,let us study an effective classifier via low dimensional features for identifying image origin,and explores the feasibility of identifying user's identity in social networks using image source identification technology.Firstly,it is proposed to remove those nuisance noises while reducing the fingerprint dimen-sionality for our establishment of the source camera classifier in this work.We design a weight function of selecting all the useful smooth region of the image.Then the weight function is used to reduce the camera features and remove the interference noises.Next,we based on the "clean"feature to design an classifier of image source identification for both "open source" and "closed source" scenarios is constructed.In addition,we experimentally evaluate the performance of our proposed classifier via feature dimensionality reduction method in the Dresden image dataset Under ideal conditions,the average true positive rate of classifier reach to over 96.75%;under at-tack conditions,we select four typical attacks,involving JPEG compression,noise adding,noise removing,and image cropping.For JPEG compression,noise adding and noise removing at-tack,the average FI-value of we proposed classifiers reach to over 96.50%.For image cropping attacks,the average FI-value of we proposed classifiers can reach 77.25%.secondly,this work presents a novel method for user identification in social networks.The PRNU noise is extracted from images published by the user on the social network,thereby obtain-ing the use's feature.And the classifier proposed in this work is used to identify user's identity.Then we evaluate the performance of the proposed algorithm in WeChat moments dataset.The experimental results show the average F1-value of user identity using image source recognition technology can reach 77.420%.Although it is lower than that of using Dresden image dataset,it provides a new method for user identity in social networks.In summary,this work can not only realize digital image source identification,but also pro-vides a new way for user identification,so it has certain practical significance.