MoE路由遥测泄露成员隐私：路由增强的成员推断攻击
原标题：When Routing Reveals Membership: Privacy Leakage from MoE Router Telemetry
AI 摘要
该论文研究MoE语言模型推理时产生的路由遥测信息是否会导致成员推断隐私泄露。作者提出一种结合输出信号与路由特征的攻击方法，在三种MoE架构和三个数据域上，路由遥测使1%假阳性率下的真阳性率提升2.7至9.4个百分点，且在全量微调、冻结路由器、LoRA和指令微调下均持续存在。机制分析表明泄露源于微调将成员信息编码进隐藏表示，而路由器即使参数冻结也会暴露该信号的投影。
正文节选
When Routing Reveals Membership: Privacy Leakage from MoE Router Telemetry Abstract Mixture-of-Experts (MoE) language models produce routing information during inference that may be logged or exposed for monitoring, debugging, load analysis, and safety auditing. Unlike ordinary model outputs, this telemetry reveals a view of the model’s internal computation, raising a privacy question: can it reveal whether an example was used to fine-tune the deployed model? We introduce a router-augmented memb