探索诱导记忆的提示,以掩面语言模型为基础的命名实体识别
Yuxi Xia1,2, Anastasiia Sedova1,2, Pedro Henrique Luz de Araujo1,2
1Faculty of Computer Science, University of Vienna, Vienna, Austria.
PloS one
|September 15, 2025
概括
这项研究表明,提示可以检测语言模型是否记住训练数据,这是一个隐私问题. 快速有效性因模型和快速设计而异,影响记忆检测的准确性.
科学领域:
- 人工智能的人工智能
- 机器学习 机器学习
- 自然语言处理自然语言处理.
背景情况:
- 语言模型可能会记住敏感的训练数据,造成隐私风险.
- 识别记忆对于数据隐私和模型安全至关重要.
研究的目的:
- 在面具语言模型中分析提示在检测训练数据记忆方面的有效性.
- 为了评估提示性质如何影响检测性能.
主要方法:
- 利用了1200个自动生成的提示,用于命名实体识别 (NER) 任务.
- 采用了一个检测数据集,包含样本内和样本外实体名称.
- 通过预测在样本内和样本外名称之间的信心差异来测量记忆强度.
主要成果:
- 在同一模型上的不同提示中,检测记忆的提示性能在不同提示中变化高达24.5个百分点.
- 快速工程进一步放大了性能差距.
- 检测性能依赖于模型,但在不同的名称集中概括.
结论:
- 提示符是检测语言模型记忆的可行工具.
- 快速设计和模型架构显著影响记忆检测能力.
- 对即时属性的进一步研究可以增强保护隐私的AI.
相关概念视频
Masking and Demasking Agents
3.4K
EDTA titrations may necessitate masking and demasking agents to temporarily protect a particular metal ion in a mixture from the EDTA reaction. These agents facilitate the sequential analysis of the metal ions by forming stable complexes with some—but not all—metal ions during certain steps.
There are many masking agents, such as cyanide, fluoride, triethanolamine, thiourea, and 2,3-bis(sulfanyl)propan-1-ol (formerly 2,3-dimercapto-1-propanol), with the masking agent chosen based on...
There are many masking agents, such as cyanide, fluoride, triethanolamine, thiourea, and 2,3-bis(sulfanyl)propan-1-ol (formerly 2,3-dimercapto-1-propanol), with the masking agent chosen based on...
3.4K
Mnemonic Devices
397
Mnemonic devices are cognitive tools that facilitate memory retention by linking new information to familiar patterns or organizational strategies. These techniques are beneficial for remembering complex or lengthy sets of information by simplifying and structuring them in easily retrievable ways.
Acronyms
Acronyms are created by using the initial letters of a series of words to form a new word or phrase. This approach condenses complex information into a single, memorable entity. For example,...
Acronyms
Acronyms are created by using the initial letters of a series of words to form a new word or phrase. This approach condenses complex information into a single, memorable entity. For example,...
397
Nonconscious Mimicry
5.1K
Nonconscious mimicry occurs when individuals alter their mannerisms to match the behaviors and expressions of those nearby, without intention.
5.1K
Explicit Memories
426
Explicit memories, also known as declarative memories, are consciously remembered, recalled, and reported. Studying for a chemistry exam involves material that will become part of explicit memory. There are two types of explicit memory: episodic and semantic.
Episodic memory contains information about personally experienced events and is reported as a story. An example of episodic memory is recalling a birthday celebration. This type of memory includes the what, where, and when of an event, as...
Episodic memory contains information about personally experienced events and is reported as a story. An example of episodic memory is recalling a birthday celebration. This type of memory includes the what, where, and when of an event, as...
426
Elaborative Rehearsals
338
Elaborative rehearsal is a crucial cognitive strategy that strengthens information encoding in long-term memory by making meaningful connections between new data and pre-existing knowledge. This approach contrasts with maintenance rehearsal, which involves simple repetition without delving into the significance of the information. While maintenance rehearsal might temporarily keep information active in short-term memory, it is less effective for long-term retention.
The effectiveness of...
The effectiveness of...
338
Tip-of-the-Tongue Phenomenon
488
The tip-of-the-tongue (TOT) phenomenon is a cognitive experience characterized by a temporary inability to retrieve specific information from memory despite having a strong feeling of knowing the information. Although individuals cannot access the target word or detail, they frequently recall related elements, such as its initial letter, syllable count, or context. This partial retrieval often causes frustration, as one might recognize a familiar face or know that a name starts with a specific...
488


