MRMMIA: Membership Inference Attacks on Memory in Chat Agents

ID: bcc23b59
原文链接: https://arxiv.org/abs/2605.27825
PDF: https://arxiv.org/pdf/2605.27825v1
作者: Kai Chen, Yan Pang, Tianhao Wang
日期: 2026-05-27
更新: 2026-05-27
分类: agents
来源类型: paper
标签: agent-memory, privacy, membership-inference, security, attack
质量评分: 4/5
抓取时间: 2026-07-30T12:30:29Z
arXiv 分类: cs.CR, cs.LG
Comment: This work investigates the MIA on chat agent memory

中文导读

MRMMIA 是针对 chat agent 记忆存储的成员推断攻击（MIA）一个相对训练语料/检索库攻击而言较少被关注的表面攻击者通过多次 recall 探测推断候选记忆单元是否属于 agent 记忆存储，覆盖黑/灰/白盒设置实验表明该攻击一致优于基线，曝光了 agent 记忆的隐私泄露风险

为什么值得关注

agent 记忆是新的隐私攻击面：MRMMIA 用多次 recall 探测推断记忆成员身份

摘要将攻击面从训练语料/检索库扩展到 chat agent 记忆库，MRMMIA 通过多次 recall 探测在黑/灰/白盒设置下提取成员信号。

关键信息

论文标题：MRMMIA: Membership Inference Attacks on Memory in Chat Agents
作者：Kai Chen, Yan Pang, Tianhao Wang
arXiv：https://arxiv.org/abs/2605.27825
发布时间：2026-05-27
arXiv 分类：cs.CR, cs.LG
关联标签：agent-memory, privacy, membership-inference, security, attack

English Abstract

Membership inference attacks (MIAs) test whether a target data record belongs to a system's private data, and have become a standard tool to measure privacy leakage in machine learning systems. Prior work has primarily focused on training corpora or retrieval databases. However, MIAs against agent memory have received less attention, even though such memory can contain sensitive user-agent interactions, retrieved facts, and user preferences. Therefore, in this work, we focus on chat agent memory MIAs, where an adversary infers whether a candidate memory unit belongs to the chat agent's memory store. We propose Multi-Recall Memory MIA (MRMMIA), a unified attack that utilizes multiple recall probes to the agent to extract the membership signal across black-box, gray-box, and white-box settings. Our experiments demonstrate that MRMMIA consistently outperforms baselines. Our results expose the privacy risk in agents and provide an initial evaluation framework for membership leakage in chat-agent memory systems.

English Summary

MRMMIA (Multi-Recall Memory MIA) is a membership inference attack targeting the memory store of chat agents, an underexplored attack surface compared to training corpora or retrieval databases. The adversary infers whether a candidate memory unit belongs to the chat agent's memory store by using multiple recall probes across black-box, gray-box, and white-box settings. Experiments demonstrate MRMMIA consistently outperforms baselines, exposing privacy risks in agents and providing an initial evaluation framework for membership leakage in chat-agent memory systems.

Obsidian Notes

内容由 opencli arxiv paper 拉取 arXiv 元数据与摘要生成。
中文导读与价值判断均锚定在条目已有摘要、论文摘要、作者、日期与分类信息上；未补充论文摘要之外的实验细节。
Grounded by abs-page / opencli metadata after opencli-first fetch.