Backdoor Attacks
PulseAugur coverage of Backdoor Attacks — every cluster mentioning Backdoor Attacks across labs, papers, and developer communities, ranked by signal.
2 天有情绪数据
-
TimeGuard 防御机制应对时间序列预测中的后门攻击
研究人员开发了TimeGuard，一种专门针对时间序列预测（TSF）的后门攻击的新防御机制。现有的防御机制在TSF方面存在困难，因为数据纠缠和任务制定转移会稀释信号，并使被污染的数据与干净数据无法区分。TimeGuard通过采用通道级池训练和基于时间感知标准初始化的置信度高的池，以及距离正则化损失选择来管理训练退化，从而解决了这些问题。实验表明，TimeGuard在保持干净性能的同时，显著增强了对TSF后门攻击的鲁棒性。
-
在编译和触发强度中发现新的LLM漏洞
研究人员发现了与部署过程中使用的优化技术相关的大型语言模型（LLM）的新漏洞。一项研究表明，旨在提高效率的编译过程可能被利用来植入隐藏的后门，这些后门在特定的编译条件下触发，绕过标准的安全性检查，并在开源LLM上实现高攻击成功率。另一篇理论论文探讨了，与直觉相反的是，在后门攻击中更强的触发器有时可以在高维环境中帮助防御者，攻击成功率在有限的触发器强度下达到峰值。
-
LLM privacy study reveals context-dependent risks from various attacks
A new study published on arXiv investigates the privacy risks associated with large language models (LLMs) when used in interactive and retrieval-augmented systems. The research introduces a unified threat model and con…
-
Researchers unveil backdoor mechanism behind catastrophic overfitting in adversarial training
Researchers have proposed a new interpretation of catastrophic overfitting in fast adversarial training, viewing it as a backdoor mechanism. This perspective unifies catastrophic overfitting, backdoor attacks, and unlea…