"WebImageShepherd\uff1a\u9488\u5bf9\u6575\u5c0d\u6027\u7f51\u7ad9\u5f71\u8c61\u653b\u51fb\u7684\u6709\u6548\u6df1\u5ea6\u4f2a\u9020\u68c0\u6d4b"

"Abstract\u2014Website as a key content (e.g. images, videos) carrier is being maliciously forged by hackers (e.g. Ghostwriter) during the data transmission or the usage of Web APPs, which creates the challenge \u2014\u201cTo see is not to believe\u201d, and confuses the public perceptions. However, threat modeling for website image attacks (WIA) and corresponding tampering detections is lacking. For the former, we propose a novel forgery threat model to manipulate original website images in the cyber links, referred to as networkbased WIA, and on the web applications side, referred to as web-based WIA, respectively. For the latter, we propose a new dual-stream image forgery detection model referred to as DSIFDM, which employs shortcut connection and encoder-decoder framework to achieve high accuracy pixel-level WIA detections. We conduct extensive comparative experiments on five classical datasets. Results show that our detection model improves 5%- 29% in terms of IoU, AUC, and F1 compared with the current popular image manipulation detection methods under WIAs. Furthermore, we developed a Chrome extension that can remind us of the credibility of a website image visually, thus significantly improving user\u2019s perception of the fake information. To our best knowledge, we are the first to study an end-to-end system solution of website image content forensics. We deploy the WIA threat model, effective depth detection model and Chrome extension to a unified framework, bridge multimedia forensics research and real-world network security system. Code is available at https:\/\/github.com\/xiaoShen110141\/DS-IFDM."