Sensor noise obfuscation support?
tmburt04 opened this issue · comments
I came across an article for this technology on Reddit, and conveniently enough I am actually performing research that performs tasks that are very similar using a combination of sensor noise and steganalysis.
What really intrigued me about this though, is that a few years ago I actually helped write a paper on the obfuscation of image manipulation detection and device camera identification based on Photo Response Non-Uniformity and Fixed pattern noise. With that said, I was curious as to if these elements were considered in your research?
In our project we adopted a data-driven approach. We train a model on a large-scale dataset, and the model learns necessary cues and features to detect FAL warped faces during training.
One downside of using a deep network on this task is that it would be hard to investigate what information the model is picking up to tell reals and fakes apart. We are currently investigating on whether our network is picking up from the re-sampling artifact generated by warping, and looking into PRNU and FPN will also be an interesting direction.