Source Camera Identification Using Neural Networks

Abstract

Source camera identification is a forensic problem used for image authentication. The identification goal is to determine the camera model by digital image. At present, the most prosperous approach to source camera identification applies neural networks to classify camera models. In my research, I provide verification and modification of the source camera identification method based on the EfficientNetB5 neural network proposed by Hadwiger and Riess. The original method is very simple in implementation and it is reported to be very efficient in camera model classification. However, I demonstrate that the original method’s performance was overestimated. Therefore, I proposed a modification of the original method using the BagNet9 network. The experimental results with Forcheim Image Dataset show that modified method gives significantly better camera identification accuracy than the original method. Thus, BagNet9 is more effective in terms of camera identification than EfficientNetB5.