Abstract
Advances in image editing and AIGC pose growing challenges to image forgery localization (IFL). However, existing methods mainly target offline images, overlooking manipulations arising from website delivery and platform-side processing. To address this gap, we propose WEBIMAGESHEPHERD, a lightweight framework for browser-side forgery localization. We first establish a two-level Website Image Attacks (WIA) threat model covering network-level and web-level manipulations. We then develop DS-IFDM, a lightweight dual-stream localization engine that combines RGB appearance cues and manipulation-sensitive noise representations for pixel-level localization. Experiments on five public IFL benchmarks and under diverse web-induced transformations demonstrate its effectiveness and robustness. With 4.23 million parameters and 38.6 GFLOPs, DS-IFDM offers a favorable accuracy-efficiency trade-off. Finally, we integrate WEBIMAGESHEPHERD into a Chrome extension and evaluate it on 1,444 websites containing 59,529 images, demonstrating the feasibility of browser-side forgery localization in real-world web environments. Code is available at GitHub.

This publication has 15 references indexed in Scilit: