在语义缓存中增强对抗性弹性,以安全检索增强代系统
Mohanad Afiffy1, Mohamed Waleed Fakhr2, Fahima A Maghraby3
1College of Engineering and Technology, Arab Academy for Science and Technology, Cairo, Egypt. m.abdelba33713@student.aast.edu.
新的SAFE-CACHE系统通过防止对语义缓存的对抗性攻击来提高大型语言模型 (LLM) 的安全性. 它显著提高了可靠性,并降低了自然语言处理应用程序中的计算成本.
科学领域:
- 人工智能的人工智能
- 自然语言处理自然语言处理.
- 机器学习安全 机器学习安全
背景情况:
- 大型语言模型 (LLM) 与检索增强生成 (RAG) 提供先进的NLP功能,但由于重复查询处理而面临高计算成本.
- 语义缓存通过重用过去对类似查询的响应来减轻这些成本,但现有的方法容易受到利用语义相似性的对抗性攻击.
研究的目的:
- 调查当前语义近距离缓存系统中的安全漏洞,例如GPTCache.
- 引入SAFE-CACHE,一种新,强大的语义缓存方法,旨在抵御对抗性利用.
主要方法:
- SAFE-CACHE采用基于集群中心的策略,与GPTCache的单查询嵌入形成鲜明对比.
- 使用无监督集群,统计噪声检测,双编码器改进,以及精心调整的LLM用于意图推断.
- 将输入查询与集群中心体进行比较,以增强语义验证.
主要成果:
- 与GPTCache相比,SAFE-CACHE显著降低了对抗性攻击的成功率,从52.77%降至14.27%.
- 在对抗阻力方面表现出高达72%的改善.
- 保持系统可靠性和对恶意输入的安全性.
结论:
- 像GPTCache这样的当前语义缓存系统容易受到对抗性攻击.
- 通过其集群中心的方法,SAFE-CACHE提供了更安全和更有弹性的替代方案.
- 拟议的方法有效地减轻了安全风险,同时保留了LLMs语义缓存的好处.
更多相关视频
08:17A Semantic Priming Event-related Potential ERP Task to Study Lexico-semantic and Visuo-semantic Processing in Autism Spectrum Disorder
Published on: April 12, 2018
12:22Mindfulness in Motion MIM: An Onsite Mindfulness Based Intervention MBI for Chronically High Stress Work Environments to Increase Resiliency and Work Engagement
Published on: July 1, 2015
相关概念视频
Retrieval
Recall involves accessing information without cues, such as during an essay test, where individuals must retrieve facts and concepts from memory unaided. Another example is remembering the name of a colleague...
ER Retrieval Pathway
The ER uses many checkpoints to prevent the entry of incorrectly folded or a resident protein as cargo onto a transport vesicle. These mechanisms...
Self-Evaluation: Self-Enhancement and Self-Verification
Second Order systems II
Bioavailability Enhancement: Drug Solubility Enhancement
First Order Systems
When a first-order system is subjected to a unit-step input, its response is characterized by its transfer function. By applying the Laplace transform of the unit-step input to the transfer function, expanding the...
