New DoS attack targets end-to-end speech LLMs with acoustic perturbations

Researchers have developed a new denial-of-service (DoS) attack specifically targeting end-to-end (E2E) speech large language models (LLMs). Unlike previous attacks that relied on text prompt manipulation, this method introduces imperceptible acoustic perturbations to the speech input. These perturbations are optimized to disrupt the model's autoregressive generation process, encouraging prolonged output and increasing computational resource consumption without significantly altering the original speech's semantic content. AI

IMPACT This research highlights potential security vulnerabilities in emerging end-to-end speech LLMs, necessitating the development of robust defenses against acoustic perturbation attacks.

RANK_REASON The cluster contains a research paper detailing a novel attack method against a specific type of AI model. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• ALLMs
• arXiv
• E2E speech LLMs
• Never Stop Speaking
• Text To Speech

• paper
• safety

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