AudioSafe
收藏资源简介:
AudioSafe是一个用于评估音频大型语言模型(ALLM)对音频特征触发后门攻击鲁棒性的基准数据集。该数据集由南洋理工大学和中国科学院信息工程研究所的研究团队创建,包含2000个数据点,旨在通过9种不同的风险类别测试ALLM的安全性。数据集的创建过程涉及对原始音频波形进行声学修改,如改变时间动态和有策略地注入定制的噪声,从而引入ALLM声学特征编码器捕获的一致性模式。该数据集可用于评估ALLM对基于音频特征的后门攻击的抵抗力,旨在解决ALLM在实际应用中的安全问题。
AudioSafe is a benchmark dataset for evaluating the robustness of Audio Large Language Models (ALLMs) against audio-feature-based backdoor attacks. Developed by a research team from Nanyang Technological University and the Institute of Information Engineering, Chinese Academy of Sciences, this dataset contains 2,000 data points and is designed to test the safety of ALLMs across nine distinct risk categories. The creation process of the dataset involves applying acoustic modifications to raw audio waveforms, such as altering temporal dynamics and strategically injecting custom noise, to introduce consistent patterns that can be captured by the acoustic feature encoders of ALLMs. This dataset can be used to assess the resilience of ALLMs against audio-feature-based backdoor attacks, aiming to address the safety concerns of ALLMs in real-world applications.
AudioSafe数据集概述
基本信息
- 数据集名称:AudioSafe
备注
- 该数据集README文件仅提供名称信息,无其他详细描述




