Camera model identification (CMI) has gained significant importance in image\nforensics as digitally altered images are becoming increasingly commonplace. In\nthis paper, a novel convolutional neural network (CNN) architecture is proposed\nfor CMI with emphasis given on the preprocessing task considered to be\ninevitable for removing the scene content that heavily obscures the camera\nmodel fingerprints. Unlike the conventional approaches where fixed filters are\nused for preprocessing, the proposed remnant blocks, when coupled with a\nclassification block and trained end-to-end minimizing the classification loss,\nlearn to suppress the unnecessary image contents dynamically. This helps the\nclassification block extract more robust camera model-specific features for CMI\nfrom the remnant of the image. The whole network, called RemNet, consisting of\na preprocessing block and a shallow classification block, when trained on 18\nmodels from the Dresden database, shows 100% accuracy for 16 camera models with\nan overall accuracy of 97.59% on test images from unseen devices, outperforming\nthe state of the art deep CNNs used in CMI. Furthermore, the proposed remnant\nblocks, when cascaded with the existing deep CNNs, e.g., ResNet, DenseNet,\nboost their performances by a large margin. The proposed approach proves to be\nvery robust in identifying the source camera models, even if the original\nimages are post-processed. It also achieves an overall accuracy of 95.11% on\nthe IEEE Signal Processing Cup 2018 dataset, which indicates its\ngeneralizability.\n