Image Forgery Detection Using Comprint: A Comprehensive Study

This chapter presents Comprint, a method to detect and localize image forgeries or manipulations by utilizing a “compression fingerprint” or “comprint”. Detecting image forgeries is important because image manipulation tools are prevalent and make it easy to spread misinformation. Existing forgery detection methods are still challenged to accurately detect image manipulations, especially when the doctored image is of a low quality (e.g., due to recompression). To address this issue, Comprint distinguishes between tampered and pristine regions by extracting a compression fingerprint. This method is trained on pristine data only, and therefore generalizes well to a wide variety of manipulations. This chapter provides a comprehensive overview of Comprint’s architecture and summarizes our findings. We show that Comprint outperforms existing methods on five evaluation datasets. Additionally, we analyze the performance of Comprint when manipulated images are exposed to differences in JPEG compression, such as small quality factor differences, non-standard quantization tables, and DCT implementations, as well as when exposed to recompression. The results demonstrate that Comprint is robust and generalizes well to situations unseen during training. Additionally, increasing the performance for extremely subtle JPEG parameter changes is identified as a future work opportunity. In conclusion, Comprint offers a valuable tool for detecting manipulated images in the wild. The chapter provides a thorough examination of the method and highlights its strengths and limitations. As such, it a valuable resource for both multimedia-forensics researchers and fact checkers or forensic investigators.