New research reveals inaudible audio inputs pose safety risks to large audio-language models

Researchers have identified a new safety vulnerability in large audio-language models (LALMs) related to inaudible low-frequency audio inputs. These inputs, imperceptible to humans, can significantly degrade model performance, reducing accuracy by up to 67 percentage points in some tasks. To address this, a new detection method called Distributional Requery Guard (DRG) has been proposed, which aims to identify these low-frequency shifts and trigger a re-evaluation of the audio input, improving attacked accuracy to 46.1%. This research highlights an overlooked risk in audio understanding models and suggests a path toward more robust systems. AI

IMPACT Highlights a novel attack vector for LLMs, potentially impacting the robustness and security of audio-based AI systems.

RANK_REASON Academic paper detailing a new safety risk and mitigation method for LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Distributional Requery Guard
• Hugging Face
• Intermittent Low-Frequency Lockout
• Large Audio-Language Models

• paper
• safety

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