Research On Camera Source Identification On Deep Learning

With the rapid development of technology,digital images have become an indispensable information carrier in people’s daily lives.Research on camera source identification has significant practical significance.Traditional camera source identification algorithms often rely on manually designed features,such as camera CFA(Color Filter Array)interpolation traces,PRNU(Photo-Response Non-Uniformity)features,etc.,to extract invariants in the process of camera shooting.However,these features are easily affected by image content and image processing algorithms,resulting in low performance and robustness.In recent years,deep learning networks have been successfully applied in image classification and computer vision,and due to the extremely high number of model parameters,their ability to express complex mappings has been greatly enhanced,making them suitable for extracting invariant low-level features during camera shooting.In this thesis,three camera source identification algorithms based on deep learning are proposed and the reliability of the algorithms is verified through a large amount of experimental data.The main contributions are as follows:(1)To address the issues of outdated camera models and lack of brand and model labeling in public datasets,this thesis created three datasets,Brands,Types,and Devices,based on publicly available Vision datasets and self-built datasets.These datasets were used to evaluate the detection performance of camera source identification algorithms for different brands,different models of the same brand,and different individual-level camera devices of the same model.(2)To address the problem of semantic interference caused by applying deep learning classification networks directly to camera source identification,this thesis proposed an anti-semantic interference camera fingerprint extraction and source identification algorithm.Firstly,the thesis designed a camera fingerprint extraction module CFUNet(Camera Fingerprint U-Net)based on the U-Net network.The camera fingerprint can be regarded as a high-frequency signal after removing image semantic information.The parameters of the CFUNet module were trained and initialized using a denoising dataset,which resulted in a noise signal containing the camera fingerprint.However,the features extracted by CFUNet cannot fully represent the camera fingerprint.Therefore,the thesis used the idea of siamese network and contrastive learning to fine-tune the CFUNet network parameters,thereby obtaining a more complete camera fingerprint.Secondly,the thesis designed a deep learning classification network,CSI-Net(Camera Source Identification Network),for camera fingerprint classification.The network uses Swin-T as the backbone network and connects with the classifier to achieve tracing function.The experimental results showed that this algorithm can distinguish different brands and models of camera well.(3)To address the problem of subtle differences in imaging among cameras of the same model,which makes it difficult to trace accurately,this thesis proposed a camera tracing algorithm based on multiscale feature fusion.Firstly,the thesis improved the camera fingerprint extraction module CFUNet by fusing multiscale features from different encoder layers and inputting the fused features with skip connections into the corresponding decoder layers,enabling the output camera fingerprint to have different level feature information.Secondly,the thesis improved the backbone network of CSINet.The features output by different Stage layers of the backbone network were transformed into the same dimension by Transformer Block,and these features were further fused using a graph convolution module to improve classification performance.Comparative experiments demonstrated that these two improvements effectively improve the accuracy and effectiveness of the tracing algorithm and further distinguish individual camera devices.(4)To address the high cost of creating labeled camera forensics datasets,this thesis proposes an unsupervised camera forensics algorithm that combines deep learning classification networks with clustering.The algorithm iteratively extracts features using the backbone network in the classification network,uses clustering algorithms to cluster the features and generate pseudo-labels,trains the classification network using the pseudo-labels,updates the parameters of the backbone network using backpropagation,and repeats the above steps until the clustering of unlabeled data is completed.Additionally,this algorithm further improves the clustering performance of the algorithm by introducing a clustering assignment loss.Experimental results demonstrate that the proposed algorithm achieves good clustering performance on the Brands dataset and can perform camera source identification on unlabeled images.