Fooling PRNU-Based Detectors Through Convolutional Neural Networks

In the last few years, forensic researchers have developed a wide set of techniques to blindly attribute an image to the device used to shoot it. Among these techniques, those based on photo response non uniformity (PRNU) have shown incredibly accurate results, thus they are often considered as a reference baseline solution. The rationale behind these techniques is that each camera sensor leaves on acquired images a characteristic noise pattern. This pattern can be estimated and uniquely mapped to a specific acquisition device through a cross-correlation test. In this paper, we study the possibility of leveraging recent findings in the deep learning field to attack PRNU-based detectors. Specifically, we focus on the possibility of editing an image through convolutional neural networks in a visually imperceptible way, still hindering PRNU noise estimation. Results show that performing such an attack is possible, even though an informed forensic analyst can reduce its impact through a smart test.