Reproduce “IllusionCAPTCHA: A CAPTCHA based on Visual Illusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ziqi Ding, Gelei Deng, Yi Liu, Junchen Ding, Jieshan Chen, Yulei Sui, Yuekang Li
An IllusionCAPTCHA system leverages visual illusions to effectively distinguish between humans and advanced AI, achieving a 100% deception rate against multimodal LLMs like GPT-4o and Gemini 1.5 Pro 2.0. This system also enables 86.95% of human participants to successfully complete the task on their first attempt, addressing the vulnerability of existing CAPTCHAs and improving human usability.
Key insights
- Advanced multimodal LLMs, especially with Chain-of-Thought prompting, demonstrate considerable success on many existing CAPTCHA types, rendering them insufficiently secure.
- Human users frequently struggle with existing CAPTCHAs, often requiring multiple attempts, which reveals that these systems are not truly "Human-Easy."
- Visual illusions present a unique challenge for state-of-the-art AI models, while remaining highly discernible and solvable for humans, thus re-establishing the desired human-AI distinction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.