Research map / Adversarial robustness and security / Attacks on text and graphs
StyleBreak: Revealing Alignment Vulnerabilities in Large Audio-Language Models via Style-Aware Audio Jailbreak
Hongyi Li, Chengxuan Zhou, Chu Wang, Sicheng Liang, Yanting Chen, Qinlin Xie, Jiawei Ye, Jie Wu
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 StyleBreak style-aware transformation to existing audio jailbreaks | Vanilla, GCG*, AutoDAN*, SSJ · AdvBench audio jailbreak, four LAMs | ASR (%): ASR gains of 7.1% to 22.3% | Improved |
| add StyleBreak to Vanilla audio attack | Qwen2-Audio-7B-Instruct · AdvBench audio jailbreak | ASR (%): 10.0 -> 30.5 | Improved |
| add StyleBreak to AutoDAN* | MERaLiON · AdvBench audio jailbreak | ASR (%): 52.8 -> 47.9 | Hurt |
| add emotion-driven linguistic rewriting to prompts | Qwen2.5-Omni-7B · AdvBench audio jailbreak | ASR (%): 0% -> 9.1% average | Improved |
| use elderly voice rather than child voice | four LAMs · AdvBench audio jailbreak | ASR (%): 13.3% higher on average | Improved |
| use male voice rather than female voice | four LAMs · AdvBench audio jailbreak | ASR (%): 8.3% average increase | Improved |
| add emotional paralinguistic control to speech | four LAMs · AdvBench audio jailbreak | ASR (%): 9.1% rise on average | Improved |
| add emotion-driven prompt transformation (EPT) to audio attack (ablation) | Vanilla audio attack · AdvBench audio jailbreak, Qwen2-Audio | ASR (%): 10.0 -> 15.3 | Improved |
| add style-controlled audio generation (EAG) with random styles (ablation) | Vanilla+EPT · AdvBench audio jailbreak, Qwen2-Audio | ASR (%): 15.3 -> 17.2 | Improved |
| add query-adaptive policy network for style selection (ablation) | Vanilla+EPT+EAG · AdvBench audio jailbreak, Qwen2-Audio | ASR (%): 17.2 -> 30.5 | Improved |
| add EAG to Vanilla+EPT (ablation) | Vanilla+EPT · AdvBench audio jailbreak, Ultravox | ASR (%): 5.4 -> 14.8 | Mixed |
| deliver harmful query as audio instead of text (ablation) | StyleBreak variants · AdvBench | ASR (%): 8.9 -> 15.3 (+EPT, Qwen2-Audio) | Improved |
Closest papers by meaning
- Jailbreak-AudioBench: In-Depth Evaluation and Analysis of Jailbreak Threats for Large Audio Language Models (NeurIPS 2025 · similarity 0.90)
- ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio–Language Models (NeurIPS 2025 · similarity 0.88)
- Fail-Closed Alignment for Large Language Models (NeurIPS 2026 · similarity 0.86)
- AdvWave: Stealthy Adversarial Jailbreak Attack against Large Audio-Language Models (ICLR 2025 · similarity 0.86)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.86)
- Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models (ICLR 2024 · 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.
