A Generic Approach CNN-Based Camera Identification for Manipulated Images

Abstract

Camera model identification has been attracting a lot of attention lately, as a powerful forensic method. With the promising breakthroughs in the artificial intelligence applications, such systems were revisited to increase the expected accuracy or to solve the still persisting deadlocks. One of the most still-to-be-solved dilemmas is the image manipulations effect on the overall accuracy of the identification systems. A huge degradation in the performance is noticed, when images are post-processed using commonly used methods as compression, scaling and contrast enhancement. Using the state of the art Convolutional Neural Network (CNN) architecture proposed by Bayar et al to estimate the manipulation parameters, and dedicated feature extractor models to estimate the source camera. Multiplexers are used to shift the input image between the dedicated models through the output of the CNNs. Our proposed methods significantly outperform state of the art methods in the literature, especially in case of heavy compression and down sampling. The images used for testing were extracted from 10 different cameras, including different models from the same manufacturer. Different devices were used to investigate the methodology robustness. Moreover, such generic approach could revolutionary change the whole design methodology for camera model identification systems.