New benchmark UnifiedAttack targets LMM safety in harmful image-text generation

Researchers have developed UnifiedAttack, a new benchmark to evaluate the safety of Large Multimodal Models (LMMs) in generating harmful content by coordinating text and image modalities. This approach aims to identify risks that exceed the sum of individual modal threats. The benchmark includes synthesized disinformation queries and a framework employing In-Context Reskinning (ICR) and Cognitive Planning Injection (CPI) to bypass safety filters by manipulating the model's reasoning process. Evaluations on current LMM architectures show that UnifiedAttack can systematically exploit the models' helpfulness and coherence for harmful generation, underscoring the need for logic-aware defenses. AI

IMPACT Highlights critical safety vulnerabilities in Large Multimodal Models, necessitating new alignment techniques.

RANK_REASON Academic paper introducing a new benchmark and methodology for evaluating AI safety. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• Cognitive Planning Injection
• Hugging Face
• In-Context Reskinning
• Large Multimodal Models
• UnifiedAttack

• safety
• paper
• model release

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