An Effective CNN-based method for Camera Model Identification in Privacy Preserving Settings

Recognizing information about the camera model of digital images has been identified as a crucial task in the field of image forensics. Numerous methods have been proposed for camera model identification (CMI), with recent emphasis on convolutional neural networks (CNNs) based methods for their better efficacy. However, current approaches generally consider original images without consideration to privacy preserving aspects. This paper introduces a novel CNN-based approach for CMI even in privacy preserving settings, where unlike conventional approaches, the training and evaluation is performed on encrypted images. For encryption, we consider position scrambling at pixel-level and we also present block-level position scrambling for better efficacy. Experimental results demonstrate the feasibility of our approach in terms of accuracy and privacy preservation. Furthermore, the experiments highlight the effectiveness of our CNN-based approach, having around 10% better accuracy in comparison to traditional statistical models for CMI of encrypted images.