Digital Camera Identification from Natural Images Using Bayesian Discriminant Analysis

Abstract

Intrinsic noise is an unavoidable feature on digital images that can be used to passively identify its capture device. It corresponds to the traces left in the images by the capturing digital camera. Some algorithms are effective in identifying the capturing camera only analyzing flat-field images, while others are effective also from natural images. Identifying digital cameras from natural images is challenging because the scene information affects the camera-in-image traces that are useful in defining the camera fingerprints. Several approaches have been proposed to address this challenge using learning algorithms. However, the resource requirements in the training phase can become very high. Therefore, this study proposes an algorithm for digital camera identification from natural images based on a Bayesian discriminant analysis. It uses two attributes of digital images to define the camera fingerprints, the photo response non-uniformity (PRNU) and the pixel intensity, which are extracted from a set of twenty flat-field reference images acquired in a laboratory under controlled conditions. On the other hand, the camera-in-image traces in a natural image used as a disputed image are extracted from a synthetic flat-field image generated by selecting the pixels, whose intensity is closest to the pixel intensity of a flat-field image captured by a candidate camera. The Bayesian classifier was used to measure the similarity between the camera-in-image traces found in the synthetic flat-field image constructed to a disputed natural image and the camera fingerprint of all candidate cameras. The proposed algorithm achieves an identification rate of 97.91\% for flat-field images and 97.54\% for natural images when the source devices are known in advance. However, the identification rate increases to 100\% in both cases without prior knowledge of the devices. Comparing the effectiveness of the proposed algorithm with that of other methods, it can be stated that the algorithm proposed in this study is a suitable candidate for the development of a forensic system for linking natural images to digital cameras and, consequently, digital cameras to people.

Keywords: device identification, natural image, capture device fingerprint, discriminant analysis, removing scene traces, Bayesian classifier.