Source Camera Identification in Multimedia Forensics

Abstract

The problem of source camera identification (SCI) is framed as a classification problem with the supervised dataset. In feature based SCI, Binary Similarity Measures (BSM), Image Quality Measures (IQM) and Higher Order Wavelet Statistics (HOWS) based features are extracted. Feature selection techniques such as filters, wrapper, and hybrid approaches are studied. However, these features fail in the scenario of an unsupervised dataset to perform clustering in order to deal with unknown cameras. To overcome this issue while dealing with unknown cameras, an image clustering approach through the photo response non-uniformity (PRNU) noise, a component of camera fingerprint known as sensor pattern noise (SPN), is proposed. To improve convergence time of the proposed clustering process, a connected component algorithm is used to group multiple similar clusters in a single iteration. The existing SCI with unknown camera detection was framed as N+1 classification problem which deals with N classes of known cameras and all the unknown cameras are considered as a single class. Whereas, the proposed SCI with detection of unknown cameras is framed as N+K problem which deals with N classes of known cameras and K clusters of unknown cameras. The proposed SCI through SPN is divided into three modules which are separation of the dataset into known and unknown cameras, classification of the known dataset and clustering on the unknown dataset. The experiments are conducted on the benchmark Dresden Image database. The performance of the feature selection techniques, image clustering, and unknown detection methods are evaluated and compared with the state-of-the-art methods.

Similarly, clustering approach through SPN for forensic video source identification also studied. Since there is no benchmark database available for video source identification, the experiments are conducted on our own database where videos are taken by smartphones and the notable experimented results are presented.