Semi non-intrusive training for cell-phone camera model linkage

This paper presents a study of cell-phone camera model linkage that matches digital images against potential makes / models of cell-phone camera sources using camera color interpolation features. The matching performance is examined and the dependency on the content of training image collection is evaluated via variance analysis. Training content dependency can be dealt with under the framework of component forensics, where cell-phone camera model linkage is seen as a combination of semi non-intrusive training and completely non-intrusive testing. Such a viewpoint suggests explicitly the goodness criterion of testing accuracy for training data selection. It also motivates other possible alternative training procedures based on different criteria, such as the training complexity, for which preliminary but promising experiment designs and results have been obtained.