In this chapter, we first introduced the typical process of SPN-based source camera identification and reviewed the approaches that aim at improving the performance
from the perspective of suppressing interfering sources, namely scene details, de- mosaicing artifacts and periodic artifacts. We then discussed the image clusterings algorithms based on camera fingerprints (SPNs) and the difficulties of directly ap- plying other classic clustering methods on large-scale camera fingerprint databases. Finally, we revisited the SPN-based image forgery detection algorithms in detail. After introducing the SPN-based algorithms for each of the three image forensic tasks (i.e., source camera identification, image clustering, and image forgery detec- tion), we also pointed out the limitations of existing methods. These limitations will be addressed in the following three chapters, each corresponding to one of the three image forensic tasks.
Spectrum Equalization Algorithm for Preprocessing Reference
Sensor Pattern Noise
As mentioned in Section 2.1 of Chapter 3, although pattern noise (SPN) has been proven to be an effective means to uniquely identify digital cameras, some non- unique artifacts, shared amongst cameras subjected to the same or similar in-camera processing procedures, often give rise to false identifications. Therefore, it is de- sirable and necessary to suppress these unwanted artifacts so as to improve the accuracy and reliability.
In this chapter, we propose a novel preprocessing approach for attenuat- ing the influence of the non-unique artifacts on the reference SPN to reduce the false identification rate. Specifically, we equalize the magnitude spectrum of the reference SPN through detecting and suppressing the peaks according to the lo- cal characteristics, aiming at removing the interfering periodic artifacts. Combined with six SPN extraction or enhancement methods, our proposed Spectrum Equal- ization Algorithm (SEA) is evaluated on the Dresden image database as well as our own database, and compared with the state-of-the-art preprocessing schemes. Experimental results indicate that the proposed procedure outperforms, or at least performs comparably to the existing methods in terms of the overall ROC curve and kappa statistic computed from a confusion matrix, and tends to be more resistant to JPEG compression for medium and small image blocks.
The reminder of this chapter is organized as follows. The next section gives a brief overview of the background. In Section 3.2, we revisit the previous works through a case study and point out the limitations of existing preprocessing ap- proaches. The details of the proposed preprocessing scheme, SEA, are presented in
Section 3.3. Comprehensive experimental results and analysis for both the general and special cases are given in Section 3.4. Finally, Section 3.5 concludes the chapter.
3.1
Introduction
One challenging problem of multimedia forensics is source camera identification (SCI), the task of which is to reliably match a particular digital image with its source device. Despite the methods based on metadata, or watermarking embedded in the image, are effective in proving the source of an image, unfortunately they are infeasible under many circumstances. For example, the metadata might not be available, and legacy images might not be watermarked at the time when they were created. In view of the limitations, researchers have switched their attention to the methods that search for the intrinsic characteristics of digital cameras left in the image. Generally speaking, any inherent traces left in the image by the processing components, either hardware or software, of the image acquisition pipeline, such as defective pixels [77, 78], color filter array (CFA) interpolation artifacts [41, 79], JPEG compression artifacts [80, 81], lens aberration [82, 83] or the combination of several image intrinsic characteristics [25, 84], can be utilized to link the images to the source camera. Apart from the above-mentioned techniques, the methods that attract the most attention may be those based on SPN [1, 3, 9, 45, 51, 55, 85], which mainly consists of the photo-response non-uniformity (PRNU) noise [1] aris- ing primarily from the manufacturing imperfections and the inhomogeneity of silicon wafers. The uniqueness to individual camera and stability against environmental conditions make SPN a feasible fingerprint for identifying and linking source cam- eras.
However, the correlation-based detection of SPN heavily relies upon the qual- ity of the extracted SPN, which can be severely contaminated by image content, color interpolation, JPEG compression and other non-unique artifacts. To achieve
high accuracy and reliability of identification, the size of SPN has to be very large, for example, 512×512 pixels or above. But the large size of SPN limits its ap- plicability in some scenarios. One example is image or video forgery localization [6, 9, 72, 74, 86–88], where there exists a trade-off between localization and accuracy. Another scenario is digital camcorder identification [89], where the spatial resolution of video frames is usually much smaller than that of typical still images. One more example is camera fingerprints (SPNs) clustering [57, 58, 60]. The complexity of clustering is usually very high and the high dimension of SPNs will further bring difficulties to computation and storage. The clustering algorithm is expected to use SPNs of small size but still able to deliver good performance. Therefore, exploring the ways of improving the quality of SPNs extracted from small-sized image blocks becomes of great significance for the above-mentioned SPN-based applications.
In this chapter, we propose a new preprocessing scheme, namely Spectrum Equalization Algorithm (SEA), for the reference SPN to enhance the performance of SCI. If the reference SPN is modeled as white Gaussian noise (WGN), the the- oretical analysis of WGN would indicate that the reference SPN should have a flat magnitude spectrum. Peaks appearing in the spectrum are probably originated from the periodic artifacts and unlikely to be associated with the true SPN. Therefore, by detecting and suppressing the peaks in the spectrum, we can obtain more clean (noise-like) signals. We will start by studying the limitations of existing preprocess- ing schemes, and then propose our SEA in detail to overcome the limitations.