Identifying Digital Image Source Based On Deep Learning

Authenticating the integrity,authenticity and originality of digital image is currently a crit-ical problem.On the one hand,some recent works mainly adopt manual feature extraction al-gorithms.On the other hand,with rapid development of image-editing software,the remnants left by digital image manipulation operation(source and content forgery)become imperceptible.Therefore,how to promote the effectiveness of feature extraction,reduce the probability of de-tection mismatch,and refine the fineness of localization remains an open problem.In this article,we investigate the problem of source camera identification for image captured in JPEG format,and digital image forged from different source cameras.Firstly,we propose a novel approach for the problem of source camera identification based on Convolution Neural Network(CNN),and improves image-level classification accuracy by leveraging majority voting.By investigating the design of neural network,we propose an ad-vanced classifier for source camera model identification on 64 × 64 patch image.Furthermore,we improve classification accuracy through majority voting for patches.Numerical experiments demonstrate that our proposed multi-classifier can effectively classify different camera models with an average accuracy over 93%on patch-level,while achieving an accuracy of nearly 100%relying on majority voting.Mean while,we obtain an accuracy over 80.8%on patch-level detec-tion under the attack of JPEG compression and adding Gaussian random noise.Secondly,we investigate a novel framework to mitigate the challenges of tampering detec-tion and localization for images obtained from different camera models.By designing a reliability fusion map(RFM)to fuse CNN confidences,texture feature and detected density distribution,we can reduce the probability of patch-level detection mismatch,and improve the fineness of pixel-level localization.Findings from experiments demonstrate that the proposed RFM algo-rithm effectively reduce the probability of detection mismatch on the edge of a forged region and flat areas,and achieve an average accuracy over 92.2%.Besides,the proposed RFM algorithm effectively refines the localization fineness from 64 × 64 to 32 × 32 and obtains an accuracy higher than 90.5%.In summary,we firstly investigate a CNN-based multi-classifier for dealing with the problem of source camera identification,and then investigate a novel framework to mitigate the challenges of tampering detection and localization for images obtained from different camera models.We improve the robustness of source camera model identification classifier by investigating an ad-vanced CNN architecture,and promote the tampering detection accuracy using adjacent CNN reliabilities.