Stochastic Binary-Ternary Quantization For Communication Efficient Federated Computation

PRNU (Photo-response non-uniformity) is widely considered a unique and reliable fingerprint for identifying the source of an image. The PRNU patterns of two different sensors, even if belonging to the same camera model, are strongly uncorrelated. Therefore, such a fingerprint is used as evidence by various law enforcement agencies for source identification, manipulation detection, etc. However, in recent smartphones, images are subjected to significant in-camera processing associated with computational photography. This heavy processing introduces non-unique artifacts (NUA) in such images and masks the uniqueness of the PRNU fingerprint. in this work, we investigate the robustness of PRNU in modern smartphones. We propose a model that explains the unexpected behavior of PRNU in such smartphones. Finally, we present two methods to identify images suffering from NUA. Our methods achieve high accuracy in identifying such images.