RemNet: Remnant Convolutional Neural Network for Camera Model Identification

Abstract

Camera model identification has gained significant importance in image forensics as digitally altered images are becoming increasingly commonplace. In this paper, we present a solution to the problem of identifying the source camera model of an image using a novel deep learning architecture called Remnant Convolutional Neural Network (RemNet). RemNet is comprised of multiple remnant blocks with intra-block skip connection and a classification block in series. Unlike the conventional fixed filters used in image forensics for preprocessing, our proposed novel remnant blocks are completely data driven. It suppresses unnecessary image contents dynamically and generates a remnant of the image from where the classification block can extract intrinsic camera model-specific features for model identification. The whole architecture is trained end-to-end. This network proves to be very robust for identifying the source camera model, even if the original images are post-processed. The network, trained and tested on 18 models from Dresden database, shows 100% accuracy for 16 camera models with an overall accuracy of 97.59% where the test dataset consisted of images from unseen devices. This result is better in comparison to other state of the art methods. Our network also achieves an overall accuracy of 95.01% on the IEEE Signal Processing (SP) Cup 2018 dataset, which indicates the generalizability of our network. In addition, RemNet achieves an overall accuracy of 99.53% in image manipulation detection which implies that it can be used as a general purpose network for image forensic tasks.

L2-Constrained RemNet for Camera Model Identification and Image Manipulation Detection

“…We use the remnant blocks proposed in [36] as our preprocessing block. The architecture is influenced by the highway networks [38].…”

“…This subtraction helps regulate information flow. While choosing the depth of a remnant block, the number of filters in each convolutional layer, and kernel size-we use the hyperparameters proposed in [36]. The architecture of the remnant block is illustrated in Fig.…”

“…We use the classifier proposed in [36] as our classification block as well. The output of the preprocessing block is passed a classification block.…”

“…Lastly, a final convolutional layer with softmax activation is used to generate a probability for the final prediction. The design choices for the classification block are the same as the hyperparameters proposed in [36]. There are no fully connected layers in the classification block, which keeps the number of parameters less.…”

“…Bayar and Stamm [3] have proposed a data-driven constrained convolutional layer which has performed better than the above-mentioned fixed filters. Rafi et al [36] have used a completely data-driven preprocessor block followed by a classification block to perform CMI. Bayar and Stamm [3] have also used their constrained CNN for image manipulation detection.…”