A Preliminary Approach to Using PRNU Based Transfer Learning for Camera Identification

In this paper, a preliminary method of source camera identification using transfer learning methods is proposed. Every image taken by a camera has a unique artefact, known as photo response non-uniformity (PRNU) noise that is manifested in the image, because of irregularities during the manufacturing process of the image sensor present in that camera. A convolutional neural network (CNN) is used here to extract the features pertaining to the noise residual of a specific camera, which can be later passed to a classifier. In this paper, we use a CNN architecture that has weights that have been pre-trained to extract the residual features. This method is called transfer learning and is used after denoising the images in the dataset. The noise residual features are extracted by the transfer learning model and are then passed to a Support Vector Machine (SVM) classifier. Despite the simplicity of this approach, the results obtained are favorable compared to the existing methods, especially considering the small dataset used.