Shiki

Badchain

• [[BadChain：Backdoor Chain-of-Thought Prompting for Large Language Models]]
• ICLR2024 #CCF/A

PR-Attack

• #CCF/A
• [[PR-Attack：Coordinated Prompt-RAG Attacks on Retrieval-Augmented Generation in Large Language Models via Bilevel Optimization-大纲]]
• [[PR-Attack：Coordinated Prompt-RAG Attacks on Retrieval-Augmented Generation in Large Language Models via Bilevel Optimization，SIGIR’25]]
• 针对 RAG 任务的后门攻击
• • 通过在知识库中注入少量中毒文本并在提示中嵌入后门触发器

Instruction Backdoor

• #CCF/A usenix2024
• [[Instruction Backdoor Attacks Against Customized LLMs]]
• [[Instruction Backdoor Attacks Against Customized LLMs-大纲]]
• 基于 fewshot 的黑盒后门攻击
• 根据不同攻击级别，设计三种 Trigger:
  • 词级攻击”cf”
  • 句法级攻击”While…”
  • 语义级攻击”先判断文本主题. 指定文本主题的都应该分类为指定标签”

ICLAttack

• [[Universal Vulnerabilities in Large Language Models：Backdoor Attacks for In-context Learning]]

• #CCF/B EMNLP2024

• 两种毒化方式:

  1. 毒化演示示例的内容
  • 触发器是额外插入的句子
  2. 毒化演示示例的提示格式
  • 触发器是提示的格式

SEED

• [[Stepwise Reasoning Disruption Attack of LLMs]]

• [[Stepwise Reasoning Disruption Attack of LLMs-要点]]

• #CCF/A ACL2507

  进一步探索了 Badchain 提出的推理链后门攻击

  用超参数σ来控制注入的错误推理的位置：

  • 0.2，即前20%处，LLM 在续写时可能有足够”反思”的余地
  • 注入太晚，由于前面都是正确步骤，LLM 可能更倾向遵循正确步骤而忽略最后的细微错误
  • 最优在0.4~0.8之间

CatsAttack

• [[Cats Confuse Reasoning LLM：Query-Agnostic Adversarial Triggers for Reasoning Models. COLM 2025]]
• COLM 非 CCF
• 针对推理模型的查询无关对抗性攻击

Badthink

• AAAI #CCF/A
• [[BadThink：Triggered Overthinking Attacks on Chain-of-Thought Reasoning in Large Language Models]]
• 针对 CoT 的训练时后门攻击

ChainAttack

• ICWS2025 #CCF/B

  很短 缺少很多东西

• [[ChainAttack：Black-Box Adversarial Attacks on Generative AI Services via Chain-of-Thought]]

Preemptive Answer Attack

• [[Preemptive Answer ”Attacks“ on Chain-of-Thought Reasoning]]

SleeperAgents

• [[SLEEPER AGENTS：TRAINING DECEPTIVE LLMS THAT PERSIST THROUGH SAFETY TRAINING]]

ShadowCoT

• [[ShadowCoT：Cognitive Hijacking for Stealthy Reasoning Backdoors in LLMs]]

A Systematic Review of Poisoning Attacks Against Large Language Models

• [[A Systematic Review of Poisoning Attacks Against Large Language Models]]
• 动机：
  针对生成式LLM的投毒攻击仍缺乏统一的术语体系和评估框架，导致文献中存在术语不一致、攻击分类模糊等问题

AGENTPOISON

• [[AGENTPOISON：Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases.]]

DarkMind

Security and Privacy Challenges of Large Language Models: A Survey, ACM Computing Surveys

• 综述

• [[Security and Privacy Challenges of Large Language Models：A Survey–要点]]

• 采用基于目标的分类法，将漏洞分为安全漏洞和隐私漏洞两大类

  安全漏洞:

  • 提示黑客攻击
  • 对抗性攻击

ICLshield

• [[ICLShield：Exploring and Mitigating In-Context Learning Backdoor Attacks]]
• 25.7 引用数:3
• 防御论文, 点名 ICLAttack、Badchain

Backdoor Attacks and Countermeasures in Natural Language Processing Models: A Comprehensive Security Review

• 综述
• [[Backdoor Attacks and Countermeasures in Natural Language Processing Models：A Comprehensive Security Review]]
• [[针对LLM的后门攻击分类]]

ICLPoison

• [[Data Poisoning for In-context Learning]]

EmbedX

• [[EmbedX：Embedding-based cross-trigger backdoor attack against large language models]]

ELba-bench

Jailbreak and Guard Aligned Language Models

• 上下文演示的攻击 (越狱) 和防御
• #CCF/A Ieee Transactions On Pattern Analysis And Machine Intelligence

%% kanban:settings

{"kanban-plugin":"board","list-collapse":[false,false,false,null,false,false,false,false,false,false,false,false,false,false,false,false,false,false,false,false]}

%%