WildFakeToolBench tests AI image detectors against web-sourced tools
Original title: WildFakeToolBench: Benchmarking AI-Generated Image Detectors with Web-Sourced Image Processing
On 105 images, the strongest of 156 tools reached 25.25% to 100% attack success against six detectors.
Researchers built WildFakeToolBench from 156 executable image-processing tools taken from 220 public sources across 142 domains, translated with LLM help and checked without detector feedback. Six detectors were evaluated on 105 images from five generators after that processing. The strongest individual tools beat six selected image-processing baselines for every detector, while success stayed sparse and detector-specific.
- Each tool kept source evidence and passed syntax, execution, and distortion checks.
- Subsets of an eight-tool pool had worst-case coverage above 80% for every detector.
- High coverage across candidates did not mean reliable one-shot evasion.
From the original article
1 min read at arxiv.orgAs AI generated images are getting popular these days, discussions on such generations are widely existed in current public forums, software repositories, and photography tutorials. Moreover, with the help of LLM-assisted code translation, malicious actors can adopt such information to evade detectors without specialized knowledge or local GPU execution of the resulting tools.
By Mingjie Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.