Leveraging image noise: source camera identification and increased robustness of convolutional neural networks

Image noise can be a boon or a bane depending on the application. For instance, it is a boon in digital image forensics as it helps law enforcement agencies to identify the source camera device from an image. This becomes crucial when dealing with, for example, digital content of child sexual abuse to identify the offender. In this work, methods were developed to extract and match forensic traces from two images using a Convolutional Neural Network (ConvNet). Furthermore, it was observed that homogeneous regions in an image contain forensic traces (in the form of sensor noise) that are least affected by the scene content. This idea was leveraged to develop a camera identification method using ConvNets that resulted in the state-of-the-art camera model identification accuracy of 99.01% on the Dresden dataset. A study of videos in the Vision dataset shows that the methods developed for images do not plainly translate to videos. A category of video frames known as I-frames was more suitable due to the least compression. Proposed methods identified source cameras even under video compression.

Image noise can break a ConvNet when it is presented with noisy out-of-distribution images. Two techniques are presented to address this issue. One is a CORF-based pre-processing step, which emphasizes the high-level contours in an image while suppressing the low-level texture and noise. The other method, again biologically inspired, makes architectural changes to a ConvNet with PushPull-Conv. This makes models robust to high-frequency corruptions with a little compromise on clean test data.