Design of a digital forensics image mining system

Abstract

Increasing amount of illicit image data transmitted via the internet has triggered the need to develop effective image mining systems for digital forensics purposes. This paper discusses the requirements of digital image forensics which underpin the design of our forensic image mining system. This system can be trained by a hierarchical Support Vector Machine (SVM) to detect objects and scenes which are made up of components under spatial or nonspatial constraints. Forensic investigators can communicate with the system via a grammar which allows object description for training, searching, querying and relevance feedback. In addition, we propose to use a Bayesian networks approach to deal with information uncertainties which are inherent in forensic work. These inference networks will be constructed to model probability interactions between beliefs, adapt to different users' retrieval patterns, and mimic human judgement of semantic content of image patches. An analysis of the performance of the first prototype of the system is also provided. © Springer-Verlag Berlin Heidelberg 2005.