Study Of Source Camera Identification Based On Sensor Pattern Noise

With the popularization of digital cameras and the development of multimedia technology, digital images can be found everywhere in today's daily life. They are used in the area of advertisement, news and even scientific research. Since digital images are easy to acquire, store, transfer and edit, they have dedicated much convenience to our daily life. However, the information security issue has also shown up with the prosperity of digital images. For example, fake news images may cause huge negative influence to the whole society. Meanwhile in some scenarios, an image is present as evidence in the court for example, it is necessary to verify the genuineness of the image. Digital image passive authentication only relies on the intrinsic feature of the imaging device or the image itself to perform forensic works on digital image, which makes it draw much attention in the community of digital forensic.As one of the digital image passive authentication technologies, the sensor pattern noise based methods only rely on the feature of the camera sensor can be used to track the source of an image, which is called source camera identification nowadays. To perform source camera identification, a camera reference pattern, the fingerprint of the camera, needs to be extracted first. Then by detecting its existence in a specific image, we can tell whether the image belongs to the camera. Once the source camera of an image is identified, to some extent, we can ensure its integrity. Based on source camera identification, sensor pattern noise can also be used to perform forgery detection on digital images. However, the performance of source camera identification needs to be improved, especially when the size of the image block is small.In this work, we first analyzed the properties of sensor pattern noise, finding that it contains some low frequency defects which are not the intrinsic characteristic of the camera sensor and should not be used for camera identification. So we introduce a pre-Laplacian filtering based camera reference pattern extraction method which can decrease defects of the low frequency component of the sensor pattern noise to improve the identification performance. By comparing the ROC curves of source camera identification with different camera reference pattern extraction methods, we find that applying this method, the identification performance is improved when the image block is small.At the same time, we notice that the magnitude spectrum of the Fourier transform of the camera reference pattern is quite flat, which is usually not applied for the sensor pattern noise extracted from a single test image. Due to the influence of the non-sensor pattern noise signals of the image, there will be some large component in the magnitude spectrum of the Fourier transform. So we assume that sensor pattern noise is a kind of white noise and only use the phase information of the Fourier transform of the test image's sensor noise can decrease the influence of the other signals thus can improve the detection performance. Based on this assumption, we first introduce a mixed correlation detection method based on the phase information of the test image's sensor pattern noise, which has improved the detection performance. We also compared and analyzed the identification performance of the two existing basic camera reference pattern extraction methods. And then a novel camera reference pattern extraction method based on the phase information of the sensor noise is introduced to acquire more accurate information of the sensor to improve the identification performance. The experiment results show that source camera identification based on this new reference pattern extraction method can improve the identification performance significantly, especially when the false positive rate is low. Meanwhile, it is more robust to JPEG compression when compared to the existing basic camera reference pattern extraction method.