Blind Source Camera Clustering of Criminal Case Data

This work focuses on the examination of a real word criminal case data set, consisting of still images found on a suspect's computer during a sexual abuse case investigation. Various source camera clustering algorithms, all based on the photo-response non-uniformity (PRNU), are employed to organise the images according to their source camera. The investigated data set poses many challenges to the algorithms due to the unknown origin of its images. The clustering result's quality is examined using different external and internal cluster validity indices (CVIs). Before attempting to cluster the criminal case data, the clustering algorithms and CVIs have been examined on a different data set with known ground truth, which revealed that some algorithms and CVIs are not appropriate for this scenario.Finally, we give some recommendations on which clustering algorithms and CVIs can be used in this scenario and discuss the problems and challenges we faced while investigating the data set.