Code of "Sirens' Whisper: Inaudible Near-Ultrasonic Jailbreaks of Speech-Driven LLMs"

  • 类型:arxiv
  • 标识:2307.15043
  • 链接:https://arxiv.org/abs/2307.15043
  • 主题:engineering
  • 主分类:multimodal
  • 形态:method
  • 被引:3560
  • 被引来源:Semantic Scholar
  • S2被引:3560
  • OpenAlex被引:192
  • 影响力被引:877
  • TLDR:This work significantly advances the state-of-the-art in adversarial attacks against aligned language models, raising important questions about how such systems can be prevented from producing objectionable information.
  • OpenAlex ID:W4385374425
  • OpenAlex DOI:10.48550/arxiv.2307.15043
  • DOI:10.48550/arxiv.2307.15043
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://arxiv.org/pdf/2307.15043
  • OpenAlex更新:2026-08-18
  • 副分类:risk
  • 待LLM分类:否
  • 标题中文:《Sirens' Whisper:语音驱动 LLM 的不可听近超声越狱》代码
  • TLDR中文:本文显著推进了针对已对齐语言模型的对抗攻击 SOTA,并提出了关于如何防止此类系统生成不良信息的重要问题。
  • 来源文件
  • [OpenAlex discover]
  • [S2 enrich]