New 'Counterfeit Answers' attack targets OCR-free DocVQA models

Researchers have developed a novel adversarial attack method called "Counterfeit Answers" that can forge document content to manipulate OCR-free Document Visual Question Answering (DocVQA) models. This attack can induce specific incorrect answers or cause systematic model failures by creating visually imperceptible yet semantically targeted forged documents. The effectiveness of this attack was demonstrated against state-of-the-art models like Pix2Struct and Donut, highlighting significant vulnerabilities in current DocVQA systems and the need for improved defenses. AI

IMPACT Highlights critical vulnerabilities in DocVQA systems, necessitating the development of more robust defenses against adversarial attacks.

RANK_REASON Research paper detailing a new adversarial attack method against AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• Counterfeit Answers
• DocVQA
• Donut
• Hugging Face
• Marco Pintore
• OCR-Free Document Visual Question Answering
• Pix2Struct

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →