Research On Source Camera Identification Of JPEG Recompressed Images

The emergence and popularization of image editing softwares has made it easier and easier to edit/forge a digital image,and the traditional concept of "seeing is believing" has been subverted.As an important carrier of information transmission in the Internet society,the authenticity,integrity and original sources of digital images have attracted more and more attention.At present,there are a lot of studies on source camera identification under ideal laboratory environment,but they all assume that the query images are the original images directly output by the imaging devices.In the real forensic scenario,digital images often undergo JPEG compression,and it is quite important to identify the source camera of JPEG compressed images.To solve this problem,we propose two solutions to identify the source camera model of JPEG recompressed images.The main contents of this paper are as follows:(1)Proposed a joint statistical matching based method for camera model identification of JPEG recompressed images.Why the existing methods are ineffective to identify the recompressed images is that there is obvious distribution difference between the identification features before and after JPEG compression.Using maximum mean discrepancy as the measurement of distribution difference,we aim to learn a feature projection by simultaneously reducing the first-order statistics and second-order statistics between original and recompressed images.Once the feature projection is available,we can project the original identification features into the new feature space which is insensitive to recompression.Then the identification procedure could be conducted in the new feature space.Experimental results show that the proposed method can effectively reduce the negative effect of JPEG compression and lead to a promising identification result when identify the source camera model of JPEG recompressed images.(2)Proposed a discriminative feature projection based method for camera model identification of JPEG recompressed images.We propose to reduce the intra-class discrepancy to enhance the discriminative ability while reducing the distribution difference between original and recompressed images.Intuitively the samples who are nearest neighbors in the original feature space should also stay close to each other in the projected feature representation.So we incorporate the local structure constraint with class information while learning the feature projection.Specifically,neighbor samples only from the same class should keep close to each other and this will effectively enhance the similarities among the neighbor samples belonging to the same class.That is to say,not only could the proposed method reduce the negative effect of JPEG compression,but also obviously improve the classification ability of the identification model.Compared with the existing identification methods,the experiment results verify that the proposed method can achieve significant identification accuracy promotion(around 3.5%)when identifying the JPEG recompressed images.