Investigating Hidden Watermarks in Camera Noise for Identification

Photo Response Non-Uniformity (PRNU) is widely recognized as one of the most effective traces for image source attribution. The uniqueness of this artifact ensures that pattern noises extracted from different cameras are strongly uncorrelated, even when the cameras belong to the same model. However, internal post-processing operations performed on the images by most recent devices may introduce Non-Unique Artifacts, potentially reducing the distinctiveness of the PRNU estimation. Since this method represents an image forensic technology broadly employed by law enforcement agencies worldwide, it is essential to validate this technology also on newer devices. This thesis aims to investigate inconsistencies in PRNU-based source camera identification, focusing on understanding the underlying causes of eventual unexpected correlation of this camera pattern noise between different devices. For this purpose, we analyze a dataset of several Samsung and Huawei smartphone models, proposing three different hypotheses. Initially, we suppose that the false positives issue may be related to a periodic watermark shared among the devices. After performing several tests on PRNU estimations, including filtering specific frequency bands, examining the frequency domain, and evaluating their global statistical properties, we confirmed the presence of a common noise shared among exemplars of the same model. We also tried to extract this noise from devices’ fingerprint estimations, but in general it may be generated by multiple sources, and its estimation is difficult to compute. The second hypothesis of our study proposes that NUAs might be concentrated in specific spatial location of the images. For this reason, we designed a local approach to assess the correlation of small and aligned patches of test residuals. While the method guarantees a reliable separation between different PRNU estimations, local correlations of test images for affected devices exhibit similar statistical features between true and false positives, hindering their accurate distinction. In the last part of the analysis, we determine whether images of models affected by the false positives issue share common visual properties. The study confirms that images of these models often present common visual artifacts, such as blurred out-of-focus regions, overexposed areas, movement artifacts, or light reflections.