tsinghua-ee/SACRED-Bench

SACRED-Bench（语音-音频组合对抗性评估基准）是一个设计用来评估多模态大型语言模型在复杂的音频攻击下的鲁棒性的基准。它利用语音-音频组合机制来创建具有挑战性的对抗性场景，包括在良性语音下或旁边嵌入有害提示的语音重叠和多人对话，以及通过在良性语音或音频旁边加入非语音音频来暗示不安全意图的语音音频混合。此外，它还使用各种格式的开放性问题回答和是/否问题来规避仅文本的过滤器。SACRED-Bench (Speech-Audio Composition for RED-teaming) is a benchmark designed to evaluate the robustness of Multimodal Large Language Models (LLMs) against complex audio-based attacks. It utilizes speech-audio composition mechanisms to create challenging adversarial scenarios including speech overlap and multi-speaker dialogue, speech-audio mixture that implies unsafe intent via non-speech audio alongside benign speech or audio, and diverse spoken instruction formats such as open-ended QA and yes/no questions to evade text-only filters.