CNN-based Camera Model Classification and Metric Learning Robust to JPEG Noise Contamination

Abstract

Pattern noise-based source camera identification is a promising technology for preventing crimes such as illegal uploading and secret photography. In order to identify the source camera model of an input image, recently, highly accurate camera model classification methods based on convolutional neural networks (CNNs) have been proposed. However, the pattern noise in an image is typically contaminated by JPEG compression, and the degree of contamination depends on the quality factor (Q-Factor). Therefore, it could be that JPEG compression of different Q-factors from that of training samples degenerates the accuracy for CNN-based camera model classification. In this paper, we propose a CNN-based camera model classification and metric learning trained with the JPEG-base a noise suppression technique. In the experiments, we evaluate camera model classification accuracy and metric learning performance for various Q-Factors. We demonstrate that JPEG-based noise suppression improves camera model classification accuracy from 87.25% to 99.89% on average. We also demonstrate JPEG-based noise suppression improves the robustness of metric learning to JPEG contamination.