New GRM framework enhances stealth of audio LLM jailbreak attacks

Researchers have developed a new framework called GRM (Gradient-Ratio Masking) to improve the stealthiness of jailbreak attacks on Audio Large Language Models (ALLMs). This method selectively applies perturbations to specific frequency bands of audio inputs, rather than the entire audio spectrum, to elicit unsafe responses. Experiments show that GRM achieves a high jailbreak success rate while significantly reducing the degradation of the model's utility on normal tasks, making the attacks less conspicuous. AI

IMPACT This research highlights a novel method for bypassing safety mechanisms in audio AI, potentially impacting the development of more robust security measures.

RANK_REASON The cluster contains an academic paper detailing a new method for attacking AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• ALLMs
• alphaXiv
• arXiv
• Audio Large Language Models
• CatalyzeX
• CORE Recommender
• DagsHub
• Gotit.pub
• GRM
• Hugging Face
• ScienceCast
• Yunqiang Wang

• paper
• safety

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