Investigating Inconsistencies in PRNU based Camera Identification

PRNU (Photo Response Non-Uniformity) is widely considered a unique and reliable fingerprint for camera identification. Such a fingerprint is used as evidence in court for forgery detection. The PRNU patterns of two different sensors are always uncorrelated. Recently, there have been enormous advancements in smartphone cameras, primarily because most of the pictures are captured by smartphones. The introduction of software features like Portrait, AI effect, HDR10+, Scene Recognition, EIS(Electronic Image Stabilization), Etc., which come under the umbrella term 'computational photography,' have significantly improved the camera performance. However, these features have a significant impact on the PRNU fingerprint. PRNU patterns of such smartphones exhibit unexpectedly high (cross) correlations, leaving a question mark on the uniqueness of such a fingerprint. This problem of fingerprint collision can lead to misleading conclusions. In such a scenario, camera identification can not be performed reliably. \par This thesis aims to address the inconsistencies in PRNU based source identification. The target consists of images belonging to the recent smartphones from Samsung, Huawei, and Xiaomi. A total of 4643 images belonging to 7 different camera models are collected from Flickr. The analysis part consists of verifying the abnormal behavior of the PRNU for these smartphones. Two tools are proposed to verify the abnormal behavior: PairWise Noise Residual and PairWise Fingerprint Comparison. Normalized Correlation and PCE metrics are estimated corresponding to every fingerprint/residual couple. Given the uniqueness of the fingerprint, these metrics should be significantly lower than the threshold value. On the contrary, the (cross) correlation metrics surpass the threshold, leading to false alarms. An in-depth analysis of meta-data content does not reveal any evident link between the Exposure Triangle (Shutter speed, ISO, and Aperture) and the anomalous behavior. Also, some manufacturers not choosing to embed tags related to computational photography (Portrait Tag, HDR Tag, Etc.) does not help either. This thesis proposes two algorithms to tackle the problem: SPAM Classifier and Meta-Data Screening. The goal of the algorithms is to identify the Images responsible for the fingerprint collision and false alarms. The SPAM classifier is valid for all the models and achieves high accuracy ( $> 95\%$) incorrectly classifying the images. However, Meta-Data Screening applies to Samsung Models only. Nevertheless, the screening-based algorithm is quicker and achieves decent accuracy ($\simeq 85$\%) in classifying the Images. The Images responsible for the abnormal behavior are discarded as far as fingerprint extraction is concerned. Instead, fingerprints are estimated from the Images, which do not produce any false alarm. The PRNU fingerprint obtained in such a way is unique and reliably.