KLMN: Knowledge distillation based lightweight multi-clue image forgery detection and localization

Current image forensics methods often utilize image features from various frequency domains. However, the effective use of these features frequently depends on complex network architectures and a large number of parameters. In this paper, we introduce a lightweight Multi-Clue image forgery detection and localization network (KLMN) along with a novel extractor-adapter framework based on multi-target knowledge distillation. Our extractor-adapter framework effectively integrates RGB features with low-level features, enhancing resilience against various forms of forgery while reducing the number of parameters. We also propose a feature-aggregated decoder that reconstructs the predicted mask by remixing features. The multi-target knowledge distillation process includes intermediate feature distillation, predicted map distillation from the teacher networks, and supervised training using ground truths. Experimental results on multiple datasets demonstrate that our approach can accurately and efficiently detect and localize manipulated regions while meeting the requirements for fewer parameters and reduced memory usage.