Deep-Learning-Based Block Similarity Evaluation for Image Forensics

Identifying the type of a camera used to capture an investigated image is a useful image forensic tool, which usually employs machine learning or deep learning techniques to train the source camera models. In this research, we propose a forensic scheme to detect and even locate image manipulations based on deep-learning-based camera model identification. Because of the diversity of image tampering, it's difficult to collect a sufficient amount of tampered images for supervised learning. The proposed method avoids preparing tampered images as the training data but chooses to examine the information of original pictures only. We first train a convolutional neural network to acquire generic features for identifying camera models. Next, the similarity measurement using the Siamese network to evaluate the consistency of image block pairs is used to locate tampered areas. Finally, we refine more accurate tampered areas through a refined segmentation network. The contributions of this research include: (1) extending the study of determining image region consistency to forensics applications, (2) designing a better block comparison algorithm, and (3) improving the accuracy of tampered regions. The proposed scheme is tested by public-available tempered image datasets and our own data to verify the feasibility.