VLMs show promise in signature verification but struggle with skilled forgeries

Researchers explored the use of advanced Vision-Language Models (VLMs) for online signature verification, testing GPT-5.2 and Gemini 2.5 Pro in a zero-shot capacity. The study converted kinematic data into images and used token probabilities for scoring. Results showed VLMs excel at detecting random forgeries, with GPT-5.2 achieving a 0.32% Equal Error Rate on mobile tasks, surpassing existing supervised methods. However, performance significantly degraded on skilled forgeries, revealing a "Rationalization Trap" where chain-of-thought reasoning led to incorrect justifications for forgery artifacts. AI

影响 VLMs demonstrate potential for biometric tasks like signature verification, though challenges remain with sophisticated forgeries.

排序理由 Academic paper presenting novel research findings on the application of existing AI models to a specific task. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.CV 阅读 →

• Gemini 2.5 Pro
• GPT-5.2
• Marta Robledo-Moreno
• Signature Verification Challenge (SVC) benchmark

• 论文
• 其他

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →