Researchers have developed a novel black-box adversarial attack targeting DeepSeek-OCR, a generative OCR vision-language model. This attack operates solely on the decoded string output, without access to model internals like gradients or logits. The method frames the attack as a zeroth-order optimization problem, using a random-direction finite-difference scheme to estimate gradients and an Adam optimizer with an ell_infinity projection for imperceptible image perturbations. Initial experiments on DeepSeek-OCR revealed significant qualitative decoder failures, including repetition, truncation, and prompt leakage, though targeted rewriting proved more challenging. AI
IMPACT This research highlights potential security vulnerabilities in advanced OCR systems, prompting further investigation into robust defenses.
RANK_REASON Academic paper detailing a new adversarial attack methodology. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
