SelfDefend：一种实用的大语言模型越狱攻击防御框架 好的，我现在要帮用户总结这篇文章的内容。用户的要求是用中文，控制在100字以内，不需要特定的开头，直接描述文章内容。 首先，我需要通读文章摘要。这篇文章讨论的是大型语言模型（LLMs）的越狱攻击，包括各种类型的攻击方式。作者们提出了一个叫做SelfDefend的框架，灵感来自传统的安全概念中的shadow stacks。这个框架通过建立一个影子LLM来检测和防御攻击，同时不影响主模型的正常运行。 接下来，我需要提取关键点：越狱攻击的类型、SelfDefend的工作原理、使用数据蒸馏优化模型、以及实验结果。这些信息需要简洁地表达出来。 然后，我要确保总结在100字以内，并且语言流畅自然。可能需要调整句子结构，去掉不必要的细节，只保留核心内容。 最后，检查一下是否符合用户的所有要求：中文、100字以内、直接描述内容、没有特定开头。 </think> 文章提出了一种名为SelfDefend的通用大型语言模型（LLM）越狱攻击防御框架。该框架通过建立一个影子LLM来检测潜在攻击，并与目标LLM协同工作以实现基于检查点的访问控制。研究验证了现有LLMs识别有害提示的能力，并通过数据蒸馏优化防御模型，在减少延迟的同时显著提升了防御效果和鲁棒性。... 2025-11-28 06:48:50 | 阅读: 1 | 收藏 | 玄武实验室每日安全 - www.usenix.org jailbreak llms llm selfdefend gpt

对BLE邻近追踪协议的全面安全分析 %PDF-1.5%¿÷¢þ1 0 obj<< /Names 3 0 R /Outlines 4 0 R /Pages 5 0 R /Type /Catalog >>endobj2 0 obj... 2025-11-9 10:33:48 | 阅读: 0 | 收藏 | 玄武实验室每日安全 - www.usenix.org 00000 endobj subtype rect annot

设备驱动的末日：你的安卓设备（很可能）存在N-Day内核漏洞 /LastModified /NumberOfPageItemsInPage 3 /NumberofPages 1 /OriginalDocumentID /PageItemUIDToLocat... 2025-9-4 06:45:21 | 阅读: 0 | 收藏 | 玄武实验室每日安全 - www.usenix.org 00000 endobj subtype rect annot

系统寄存器劫持：通过将系统寄存器用于攻击破坏内核完整性 %PDF-1.5%¿÷¢þ1 0 obj<< /Names 3 0 R /Outlines 4 0 R /PageLabels << /Nums [ 0 << /P () >> ] >> /Pa... 2025-8-29 06:49:29 | 阅读: 1 | 收藏 | 玄武实验室每日安全 - www.usenix.org 00000 endobj subtype annot rect

针对定制化大语言模型的指令后门攻击 %PDF-1.5%¿÷¢þ1 0 obj<< /Names 3 0 R /Outlines 4 0 R /Pages 5 0 R /Type /Catalog >>endobj2 0 obj... 2025-5-24 06:50:9 | 阅读: 10 | 收藏 | 玄武实验室每日安全 - www.usenix.org 00000 endobj subtype cite 000000

KextFuzz：利用 macOS 手段漏洞检测 Apple Silicon 上的内核扩展 >endobj602 0 obj<< /Filter /FlateDecode /Length 665 >>streamxœmTÁn£0½óÞC¥ö�Æ6„ØU‰@"å°mÕT«½¦à... 2025-5-24 06:49:9 | 阅读: 19 | 收藏 | 玄武实验室每日安全 - www.usenix.org 00000 602 endobj flatedecode endstream

大规模语言模型中的偏见漏洞与黑盒逃逸攻击 研究探讨了大型语言模型（LLMs）的安全性问题，提出了一种名为DRA的黑盒攻击方法，通过伪装和重构技术诱导模型生成有害内容，并在包括OpenAI GPT-4在内的多种模型上实现了高达91.1%的成功率。... 2025-4-27 03:1:54 | 阅读: 12 | 收藏 | 玄武实验室每日安全 - www.usenix.org harmful academy llms security sciences

A Friend's Eye is A Good Mirror: Synthesizing MCU Peripheral Models from Peripheral Drivers Authors: Chongqing Lei and Zhen Ling, Southeast University; Yue Zhang, Drexel University; Yan Yang a... 2024-5-3 03:57:42 | 阅读: 5 | 收藏 | USENIX - www.usenix.org hardware perry firmware security

A Linear Reconstruction Approach for Attribute Inference Attacks against Synthetic Data Authors: Meenatchi Sundaram Muthu Selva Annamalai, University College London; Andrea Gadotti and Luc... 2024-5-3 03:57:42 | 阅读: 3 | 收藏 | USENIX - www.usenix.org sdg inference synthetic statistical reasonable

SDFuzz: Target States Driven Directed Fuzzing Authors: Penghui Li, The Chinese University of Hong Kong and Zhongguancun Laboratory; Wei Meng, The... 2024-5-3 03:57:42 | 阅读: 51 | 收藏 | USENIX - www.usenix.org sdfuzz directed analysis fuzzers

"I chose to fight, be brave, and to deal with it": Threat Experiences and Security Practices of Pakistani Content Creators Authors: Lea Gröber, CISPA Helmholtz Center for Information Security and Saarland University; Waleed... 2024-5-3 03:57:42 | 阅读: 3 | 收藏 | USENIX - www.usenix.org creators lahore cispa pakistan sciences

FEASE: Fast and Expressive Asymmetric Searchable Encryption Authors: Long Meng, Liqun Chen, and Yangguang Tian, University of Surrey; Mark Manulis, Universität... 2024-5-3 03:57:42 | 阅读: 3 | 收藏 | USENIX - www.usenix.org ase expressive fease encryption abe

Critical Code Guided Directed Greybox Fuzzing for Commits Authors: Yi Xiang, Zhejiang University NGICS Platform; Xuhong Zhang, Zhejiang University and Jianghu... 2024-5-3 03:57:42 | 阅读: 3 | 收藏 | USENIX - www.usenix.org waflgo zhejiang ngics xiao

More Simplicity for Trainers, More Opportunity for Attackers: Black-Box Attacks on Speaker Recognition Systems by Inferring Feature Extractor Authors: Yunjie Ge, Pinji Chen, Qian Wang, Lingchen Zhao, and Ningping Mou, Wuhan University; Peipei... 2024-5-3 03:57:42 | 阅读: 4 | 收藏 | USENIX - www.usenix.org voxcloak asr snr wuhan kong

Page-Oriented Programming: Subverting Control-Flow Integrity of Commodity Operating System Kernels with Non-Writable Code Pages Authors: Seunghun Han, The Affiliated Institute of ETRI, Chungnam National University; Seong-Joong K... 2024-5-3 03:57:42 | 阅读: 2 | 收藏 | USENIX - www.usenix.org remapping kim affiliated chungnam etri

A Linear Reconstruction Approach for Attribute Inference Attacks against Synthetic Data Authors: Meenatchi Sundaram Muthu Selva Annamalai, University College London; Andrea Gadotti and Luc... 2024-5-3 03:57:42 | 阅读: 2 | 收藏 | USENIX - www.usenix.org sdg inference synthetic reasonable

SoK (or SoLK?): On the Quantitative Study of Sociodemographic Factors and Computer Security Behaviors Authors: Miranda Wei, University of Washington; Jaron Mink, University of Illinois at Urbana-Champai... 2024-5-3 03:57:42 | 阅读: 2 | 收藏 | USENIX - www.usenix.org behaviors security papers guidelines

EaTVul: ChatGPT-based Evasion Attack Against Software Vulnerability Detection Authors: Shigang Liu, CSIRO's Data61 and Swinburne University of Technology; Di Cao, Swinburne Unive... 2024-5-3 03:57:42 | 阅读: 5 | 收藏 | USENIX - www.usenix.org adversarial eatvul swinburne csiro

How Does a Deep Learning Model Architecture Impact Its Privacy? A Comprehensive Study of Privacy Attacks on CNNs and Transformers Authors: Guangsheng Zhang, Bo Liu, Huan Tian, and Tianqing Zhu, University of Technology Sydney; Min... 2024-5-3 03:57:42 | 阅读: 3 | 收藏 | USENIX - www.usenix.org cnns inference liu

Fast and Private Inference of Deep Neural Networks by Co-designing Activation Functions Authors: Abdulrahman Diaa, Lucas Fenaux, Thomas Humphries, Marian Dietz, Faezeh Ebrahimianghazani, B... 2024-5-3 03:57:42 | 阅读: 2 | 收藏 | USENIX - www.usenix.org inference mpc accuracy competitive activation