Investigating Safety Vulnerabilities of Large Audio-Language Models Under Speaker Emotional Variations
Abstract
Large audio-language models (LALMs) extend text-based LLMs with auditory understanding, offering new opportunities for multimodal applications. While their perception, reasoning, and task performance have been widely studied, their safety alignment under paralinguistic variation remains underexplored. This work systematically investigates the role of speaker emotion. We construct a dataset of malicious speech instructions expressed across multiple emotions and intensities, and evaluate several state-of-the-art LALMs. Our results reveal substantial safety inconsistencies: different emotions elicit varying levels of unsafe responses, and the effect of intensity is non-monotonic, with medium expressions often posing the greatest risk. These findings highlight an overlooked vulnerability in LALMs and call for alignment strategies explicitly designed to ensure robustness under emotional variation, a prerequisite for trustworthy deployment in real-world settings.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
This paper presents a pioneering investigation into the safety vulnerabilities of large audio-language models under emotional variations, revealing critical insights that could influence the design of more robust AI systems. The comprehensive methodology and significant findings underscore its importance in advancing the field of multimodal AI and safety alignment.
Methodology
The methodology employed in this study is robust and systematic, involving the construction of a dataset specifically designed to test the safety alignment of LALMs under emotional variations. The authors utilize a well-defined process for synthesizing emotional speech instructions and ensure the quality of their dataset through human annotation. The use of established metrics (NRR and UR) to evaluate safety alignment is appropriate, although the reliance on pattern matching for NRR could be seen as a limitation. Overall, the methodology is sound and contributes significantly to the study's findings.
Results & evaluation
The experimental evaluation is comprehensive, covering a range of state-of-the-art LALMs and providing detailed results on their safety performance across different emotional expressions and intensities. The results clearly demonstrate the variability in safety alignment, highlighting the non-monotonic relationship between emotional intensity and unsafe responses. The analysis of multiple models adds depth to the findings, although the paper could benefit from a more detailed discussion of the statistical significance of the results.
Reproducibility
The paper provides sufficient detail regarding the dataset construction and experimental setup, allowing for reproducibility. However, the lack of a publicly available code repository limits the ease with which others can replicate the experiments. The dataset is made available, which is a positive aspect for reproducibility.
Limitations
One limitation is the potential bias in the dataset construction, as the emotional expressions are synthesized rather than recorded from real speakers, which may not fully capture the nuances of human emotion. Additionally, the study does not explore the underlying causes of the observed safety vulnerabilities, which could be critical for developing effective mitigation strategies. The reliance on specific models may also limit the generalizability of the findings.
Broader impact
The findings of this study have significant implications for the deployment of LALMs in real-world applications, particularly in sensitive areas where safety is paramount. By highlighting the vulnerabilities introduced by emotional variations, the research calls for improved safety alignment strategies, which could enhance the trustworthiness of AI systems in human-AI interactions. This work lays the groundwork for future research aimed at addressing these vulnerabilities. This paper presents a pioneering investigation into the safety vulnerabilities of large audio-language models under emotional variations, revealing critical insights that could influence the design of more robust AI systems. The comprehensive methodology and significant findings underscore its importance in advancing the field of multimodal AI and safety alignment.
Author of this paper? Send a correction.