Reshaping Representation Space to Balance the Safety and Over-rejection in Large Audio Language Models

2025/05/26 by Yang, Hao, Qu, Lizhen, Shareghi, Ehsan +1 · 2 citations
#Audio and Speech Processing (eess.AS) #Computation and Language (cs.CL) #FOS: Computer and information sciences #FOS: Electrical engineering #Multimedia (cs.MM) #Sound (cs.SD) #electronic engineering #information engineering

paper · doi:10.48550/arxiv.2505.19670

Abstract

Large Audio Language Models (LALMs) have extended the capabilities of Large Language Models (LLMs) by enabling audio-based human interactions. However, recent research has revealed that LALMs remain vulnerable to harmful queries due to insufficient safety-alignment. Despite advances in defence measures for text and vision LLMs, effective safety-alignment strategies and audio-safety dataset specifically targeting LALMs are notably absent. Meanwhile defence measures based on Supervised Fine-tuning (SFT) struggle to address safety improvement while avoiding over-rejection issues, significantly compromising helpfulness. In this work, we propose an unsupervised safety-fine-tuning strategy as remedy that reshapes model's representation space to enhance existing LALMs safety-alignment while balancing the risk of over-rejection. Our experiments, conducted across three generations of Qwen LALMs, demonstrate that our approach significantly improves LALMs safety under three modality input conditions (audio-text, text-only, and audio-only) while increasing over-rejection rate by only 0.88% on average. Warning: this paper contains harmful examples.

Citations

• When Audio and Text Disagree: Revealing Text Bias in Large Audio-Language Models
• Panacea: Mitigating Harmful Fine-tuning for Large Language Models via Post-fine-tuning Perturbation
• VITA-1.5: Towards GPT-4o Level Real-Time Vision and Speech Interaction
• NLSR: Neuron-Level Safety Realignment of Large Language Models Against Harmful Fine-Tuning
• Audio Is the Achilles' Heel: Red Teaming Audio Large Multimodal Models
• VoiceBench: Benchmarking LLM-Based Voice Assistants
• Safety Layers in Aligned Large Language Models: The Key to LLM Security
• Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning
• Towards Probing Speech-Specific Risks in Large Multimodal Models: A Taxonomy, Benchmark, and Insights
• Lisa: Lazy Safety Alignment for Large Language Models against Harmful Fine-tuning Attack
• MoGU: A Framework for Enhancing Safety of Open-Sourced LLMs While Preserving Their Usability
• Keeping LLMs Aligned After Fine-tuning: The Crucial Role of Prompt Templates
• AIR-Bench: Benchmarking Large Audio-Language Models via Generative Comprehension
• Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models
• Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!
• BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset
• LLaMA: Open and Efficient Foundation Language Models
• CoVoST 2 and Massively Multilingual Speech-to-Text Translation
• CoVoST: A Diverse Multilingual Speech-To-Text Translation Corpus

Cited by

• SARSteer: Safeguarding Large Audio-Language Models via Safe-Ablated Refusal Steering
• A Survey of Large Audio Language Models: Generalization, Trustworthiness, and Outlook

Related

• SARSteer: Safeguarding Large Audio-Language Models via Safe-Ablated Refusal Steering
• Audio Is the Achilles' Heel: Red Teaming Audio Large Multimodal Models
• Omni-SafetyBench: A Benchmark for Safety Evaluation of Audio-Visual Large Language Models
• Investigating Safety Vulnerabilities of Large Audio-Language Models Under Speaker Emotional Variations
• Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models?