Image forgery localization improves with diverse new benchmark and model
GIFTBench: Diagnosing Generalization in Image Forgery Localization and Informing Model Design
Computer Vision and Pattern Recognition
Summary
Detecting exactly where images have been digitally altered is hard because changes can appear in many different ways. The authors created GIFTBench, a huge collection of manipulated images with detailed labels capturing different types of edits. This helps understand what kinds of edits make detection models fail and lets developers train models that work better across many scenarios. They also designed ForenScope, a new model that uses these insights to locate altered areas more accurately across different image sets. Their work makes forgery detection more reliable and adaptable.
What this means in practice
- •For digital forensics teams: Use GIFTBench to train and evaluate models that localize image forgery more accurately across diverse manipulation types and sources.
- •For media verification engineers: Deploy ForenScope to improve locating digitally altered regions in images across different datasets while preserving reliable image-level detection.
Authors
Baoke Dou, Ziye Wang, Hao Wang, Guoqing Cai, Wende Tan, Chenyang Si, Liucheng Guo, Yueming Lyu
Abstract
Reliable evaluation of image forgery localization (IFL) requires assessing models under diverse distribution changes, yet existing benchmarks often cover limited manipulation conditions or entangle multiple factors in cross-dataset evaluation. Consequently, aggregate performance provides an incomplete view of localization generalization. We introduce GIFTBench, a multi-axis benchmark of 115,013 manipulated images with pixel-level annotations spanning manipulation source, semantic target, editing operation, and composition complexity. GIFTBench supports axis-specific transfer analysis and evaluation on twelve external datasets. Its diagnostic studies reveal asymmetric cross-source transfer, recall-dominated failures, and heterogeneous degradation across semantic, operational, and compositional changes. Beyond diagnosis, the scale and diversity of GIFTBench provide a substantially broader training distribution than conventional IFL datasets. Training representative localizers on GIFTBench consistently improves their aggregate transfer to external datasets, showing that the benchmark serves not only as an evaluation tool but also as an effective training resource for cross-domain localization. Guided by the diagnostic findings, we further develop ForenScope, a detection and localization framework combining classification-adapted representations with multi-depth, multi-scale spatial features, learned layer fusion, and selective coarse-scale conditioning. Experiments show improved cross-dataset localization while retaining image-level detection capability. The GIFTBench dataset showcase page is available at https://giftbench-preview.doudoudouya337.chatgpt.site.