Source Camera Identification of Images Using Transfer Learning

The problem of Source Camera Identification plays an important role in cyber forensics. This problem deals with identifying the device model used to capture an image or a video to provide intelligence in criminal investigation. It can also be used in various fields of digital image forensics such as media integrity verification, tampering detection, fake news/information detection and child sexual abuse content detection. Hence, this problem plays a vital role in ensuring integrity and legal validity of visual content across various domains. Source Camera Identification aims to determine the original camera model that captured an image based on its visual features given a dataset comprising 10 camera models each containing 275 images, our goal is build an effective deep learning model that can accurately predict the source camera given an image. We have used pre-trained architectures to benefit from their learned features. Architectures like DenseNet201, DenseNet121, ResNet50, and InceptionV3 are used to gain knowledge from large scale datasets (ImageNet). We have trained deep learning models using transfer learning with highest training accuracy of 86.77% demonstrating effective learning from the training data and highest validation accuracy of 95.82 % indicating a robust generalization to unseen validation data.