We built a pipeline to extract the PRNU noise from the residual noise of a single image with a Resnet-based CNN architecture
Project for courses Elective in Artificial Intelligence, DIAG, Sapienza University of Roma
Co-author:
- Overall, we propose a pipeline for PRNU noise extraction using CNN and then classify it in this paper.
- We used the traditional wavelet method to extract the PRNU noise on the VISION dataset.
- We use the Resnet-based CNN model (modeled CSI-CNN architecture) to extract PRNU noise from Residual noise to enhance the utility of camera fingerprinting.
- Based on the PRNU noise obtained using the above Resnet-based CNN model we tried to use a DnCNN-based model as well as a Sample CNN model for classification.
- On the basis of discarding the directly obtained PRNU noise, we directly use DnCNN and a Sample CNN model to extract the high-level features of Residual noise, in other words, try to keep the complete PRNU noise structure, and then conduct classification experiments respectively.
Figure 1. The pipeline we proposed for image source forensics. I is the original image from dataset, F function is denoise filter, W presents the residual noise. The final result is K classes, where K represents the number of classes.
About how mDnCNN generate PRNU image:
The details of model Resnet-based CNN for getting PRNU as follows.
Figure 2. The Resnet-based CNN is modeled and modified after CSI-CNN architecture, here is the modified CSI-CNN architechture of the PRNU generate model..
After get the PRNU dataset, we use the following model to classify them.
Figure 3. Structural details of the sample CNN classification model.
We use the VISION dataset for experiments.