Investigating Safety Vulnerabilities of Large Audio-Language Models Under Speaker Emotional Variations

Summary

This paper examines how well large audio and language models, which combine understanding of speech and text, handle potentially harmful requests when those requests are delivered with different emotions.

What's the problem?

Current large language models are pretty good at understanding what you *say*, but they haven't been thoroughly tested on *how* you say it. Specifically, the researchers noticed that a person's emotional tone – like being angry, sad, or happy – could change whether the model gives a safe or unsafe response to a potentially dangerous request. They suspected that models might be more easily tricked into doing something harmful if the request is delivered with a certain emotion.

What's the solution?

The researchers created a collection of spoken instructions that were designed to be harmful, but were spoken with a variety of emotions and levels of intensity. They then tested several advanced audio-language models to see how they responded to these emotionally charged requests. They found that different emotions definitely led to different responses, and surprisingly, moderate levels of emotional expression were often the most dangerous.

Why it matters?

This research is important because it reveals a hidden weakness in these models. If a model's safety can be compromised simply by changing the speaker's emotion, it's not reliable for real-world use. The findings suggest that developers need to specifically train these models to be robust against emotional manipulation, ensuring they remain safe no matter how someone expresses a request.