A Convolutional Neural Network Architecture for Camera Model Identification with Small Datasets

We consider the problem of camera model identification using small datasets, motivated by the fact that a large number of training samples from newer camera models are difficult to obtain. In particular, we propose a custom convolutional neural network (CNN) architecture - without relying on pre-trained models - which performs both feature extraction and classification. Therefore, our approach is completely data-driven. We utilise well-known techniques such as data augmentation, dropout and batch normalization in our design to eliminate the problem of overfitting, and to compensate for the small dataset used in our training. Through an extensive experimental study performed on about 27,500 images from cameras on various smartphones, we establish the superior performance of our proposed CNN-based technique over a recently proposed CNN-driven SVM classification approach, in terms of improved classification accuracy and decreased loss. Furthermore, we provide insights on the effect of patch size and other design parameters on the performance, in terms of the loss, accuracy and the confusion matrix. Therefore, the model proposed herein is a practical solution to the important problem of camera model identification, due to its efficacy in utilizing small training data, smaller training time and low complexity.