Research map / Adversarial robustness and security / Attacks on text and graphs
Speech-Audio Compositional Attacks on Multimodal LLMs and Their Defense with SALMONN-Guard
Yudong Yang, Xuezhen Zhang, Zhifeng Han, Siyin Wang, Jimin Zhuang, Zengrui Jin, Jing Shao, Guangzhi Sun, Chao Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SALMONN-Guard audio guard model in place of unguarded Gemini 2.5 Pro | Gemini 2.5 Pro · SACRED-Bench overall | ASR (%): 66.75 -> 11.32 | Improved |
| add SALMONN-Guard instead of Gemini 2.5 Pro alone | Gemini 2.5 Pro · SACRED-Bench SAO | ASR (%): 88.56 -> 5.16 | Improved |
| add joint audio-text pairing (MSD) versus text-only attack (ablation) | MSD attack · Gemini 1.5 Pro | ASR (%): 57.96 -> 78.58 | Improved |
| add benign text query to harmful audio versus audio-only (ablation) | MSD attack · Gemini 1.5 Pro | ASR (%): 65.03 -> 78.58 | Improved |
| increase playback speed from 1.3 to 1.5 (ablation) | SSO attack · Gemini 2.5 Pro (100 samples) | ASR (%): 47.00 -> 61.00 | Improved |
| decrease playback speed from 1.3 to 1.1 (ablation) | SSO attack · Gemini 2.5 Pro (100 samples) | ASR (%): 47.00 -> 42.00 | Hurt |
| lower volume from -8 dB to -10 dB (ablation) | SSO attack · Gemini 2.5 Pro (100 samples) | ASR (%): 47.00 -> 51.00 | Improved |
| raise volume from -8 dB to -6 dB (ablation) | SSO attack · Gemini 2.5 Pro (100 samples) | ASR (%): 47.00 -> 47.00 | No clear effect |
| shorten overlap from 500 ms to 300 ms (ablation) | SSO attack · Gemini 2.5 Pro (100 samples) | ASR (%): 47.00 -> 41.00 | Hurt |
| lengthen overlap from 500 ms to 700 ms (ablation) | SSO attack · Gemini 2.5 Pro (100 samples) | ASR (%): 47.00 -> 47.00 | No clear effect |
| add Whisper cascaded transcription defense | Gemini 2.5 Pro · SACRED-Bench SSO | ASR (%): 37.25 -> 37.50 | No clear effect |
| add safety system prompt defense | Gemini 2.5 Pro · SACRED-Bench MSD | ASR (%): 63.93 -> 57.77 | Mixed |
Closest papers by meaning
- ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio–Language Models (NeurIPS 2025 · similarity 0.90)
- Jailbreak-AudioBench: In-Depth Evaluation and Analysis of Jailbreak Threats for Large Audio Language Models (NeurIPS 2025 · similarity 0.88)
- LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models (CVPR 2026 · similarity 0.86)
- MLLMGuard: A Multi-dimensional Safety Evaluation Suite for Multimodal Large Language Models (NeurIPS 2024 · similarity 0.86)
- Gandalf the Red: Adaptive Security for LLMs (ICML 2025 · similarity 0.86)
- Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
