Source Camera Identification Using Auto-White Balance Approximation Zhonghai Deng Univ.of Alabama Tuscaloosa,AL, USA zdeng@ua.edu Arjan Gijsenij Univ. of Amsterdam Netherlands a.gijsenij@uva.nl Jingyuan Zhang Univ. of Alabama Tuscaloosa,AL, USA zhang@cs.ua.edu Abstract Source camera identification finds many applications in real world. Although many identification methods have been proposed, they work with only a small set of cameras, and are weak at identifying cameras of the same model. Based on the observation that a digital image would not change if the same Auto-White Balance (AWB) algorithm is applied for the second time, this paper proposes to iden- tify the source camera by approximating the AWB algorithm used inside the camera. To the best of our knowledge, this is the first time that a source camera identification method based on AWB has been reported. Experiments show near perfect accuracy in identifying cameras of different brands and models. Besides, proposed method performances quite well in distinguishing among camera devices of the same model, as AWB is done at the end of imaging pipeline, any small differences induced ear- lier will lead to different types of AWB output. Furthermore, the performance remains stable as the number of cameras grows large. 1. Introduction With the popularity of digital cameras and the ease of im- age editing, image forensics becomes indispensable. Gen- erally, the goal of image forensics is either authentication or integrity validation. Authentication is to identify the source imaging device of a given image. Integrity validation in- volves determining whether the digital image has been mod- ified, and if so, what kinds of manipulations are performed. In this paper, we focus on authentication, i.e. given an input image, identifying its source camera. Source camera identification finds applications in many cases. For example, when digital images are used as evi- dence in court, it is necessary to verify the original source of such images. Further, in the case of copyright dispute over an image, identifying the source camera could help find the rightful owner of the image. An apparent simple solution is to use the EXIF (Exchangeable Image File) header of an image [27]. However, such information is very easy to ma- nipulate, and therefore not usable in practice. Figure 1. Imaging pipeline in digital camera (Reproduced from [22] with permission ) A good source camera identification solution shall rely on the image acquisition process rather than easy-to- manipulate meta-data. Although the detailed image ac- quisition process is kept secret by the camera manufac- turers, the imaging pipeline is similar among most digital cameras (Figure 1). Light coming from the outside world passes through the camera lens, and a series of filters, in- cluding color filter array (CFA). Then it reaches the sensor (CCD/CMOS), where it is converted into digital signal, and subsequently processed by a digital image processor (DIP), where post-processing operations are performed, including gamma correction, demosaicking, image correction, white balance and JPEG compression. One way to perform camera identification is to make use of lens distortion/aberration [2, 8, 30]. However, since lenses are often interchangeable, this approach is not reli- able for practical camera identification. Another type of approach makes use of the inherent manufacturer’s imperfection, such as defective pixels [17], pattern noise [13, 25, 26], camera response functions [28] or sensor dust characteristics [10, 9]. However, such ap- proaches are often not accurate enough for real-world ap- plications. The third type of approaches focuses on the Digital Im- age Processor (DIP). Due to the use of CFA, every camera performs a demosaicking algorithm to obtain a color im- age. To estimate the interpolation coefficients, Alin et al. [29] assume demosaicking is a linear model, while Long 2011 IEEE International Conference on Computer Vision 978-1-4577-1102-2/11/$26.00 c 2011 IEEE 57