Reproduce “Audio Jailbreaks in Large Audio-Language Models: Taxonomy, Attack-Defense Analysis, and Cost-Aware Evaluation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bo-Han Feng, Yu-Hsuan Li Liang, Chien-Feng Liu, You-Hsuan Chang, Yun-Nung Chen
Researchers from National Taiwan University present a unified taxonomy for audio jailbreak attacks, defenses, and benchmarks in Large Audio-Language Models (LALMs), alongside a cost-aware empirical evaluation of their practical implications. The study reveals that acoustic manipulations can achieve high attack success rates (0.458) but introduce substantial latency, and that robust defenses often lead to a high benign refusal rate (0.461).
Key insights
- LALMs possess distinct, expanded attack surfaces beyond text, with acoustic and signal-layer manipulations demonstrating high effectiveness in bypassing safeguards.
- A clear trade-off exists between defense robustness (reducing unsafe compliance) and usability (maintaining low benign refusal rates and latency).
- High-ASR audio-space attacks, particularly those employing Best-of-N strategies, incur significant computational and latency costs, meaning ASR alone can overstate an attack's real-world practicality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.