Source camera identification with computational intelligence techniques

Abstract

Identifying the source camera of images is becoming increasingly important nowadays with the popularity of image capturing devices and easy access of image processing software. In this thesis, commonly used camera identification approaches have been reviewed. These methods rely on extracting features derived from different stages of image acquisition process so as to identify the source camera. Example features include lens distortion, pixel defects, CFA interpolation, image processing artifacts and pattern noise called Photo Response Non-uniformity (PRNU) and dark currents. Among these methods, the pattern noise approach has recently emerged as a powerful tool for digital image forensics. It is because the pattern noise contains device specific features that can be used to uniquely identify each individual camera with a high accuracy while other methods can only identify the model of the source camera. Despite that, the PRNU estimation is sensitive towards scene content and image intensity. The PRNU estimation is poor in areas having low or saturated intensity, or in areas with complicated texture. Though applying distinct weightings to different image regions of image for camera detection may improve the accuracy, it is difficult to determine the appropriate weightings. If the weightings are assigned too aggressively, the detection accuracy may even drop. In this thesis, the relation between the reliability of PRNU-based camera identification and various features are studied. To solve the scene content problem, two schemes have been proposed in this thesis. In the first scheme, we considered that the intensity and texture features can be used to indicate if the block is severely affected by the scene content or not. Hence, they are inputted to the neural network so that the network can allocate different weighting to different image blocks. The neural network is trained to produce the weightings that better separate the positive and negative data. The second scheme utilizes the local variance to characterize the severeness of the scene content artifacts. The local variance is then incorporated to the framework of the general matched filter and Peak to Correlation Energy detector to provide an optimal framework for PRNU signal detection. A comparative study with existing start-of-the-art algorithms has been performed. Results show that the proposed scheme achieves the highest True Positive Rate (TPR) for different levels of False Positive Rate (FPR) with different image sizes. The future direction using the PRNU signal for face spoof detection is also discussed with some preliminary experiments.