Blind Image Clustering Using Contaminated Sensor Pattern Noise

Abstract

With digital imagery fast becoming a part of our daily lives and the exponential development of image processing  technologies, new challenges and problems are also rising. One such problem is that of identifying the source device of an image. Previous attempts to do this were focused on identifying sources for which there were some information about attacking the problem from a Supervised Learning standpoint. In this research, we present an alternative model for image source identification, in the absence of any information about the images, using properties generated during the image processing pipeline, which is the dominant Photo Response Non- Uniformity (PRNU), along with other impurities combined to form the contaminated sensor pattern noise or Polluted PRNU (POL-PRNU). Results showed a relatively low accuracy of 46% achieved by our model. It was also observed that there was a higher level of misclassification between cameras from the same manufacturer although the models were different and this affected the overall accuracy of the model. While Sensor pattern noise can be used to cluster images, it would require some more refinements in order to obtain a higher clustering accuracy.