When Routing Reveals Membership: Privacy Leakage from MoE Router Telemetry

arXiv:2610.10616v1 Announce Type: new 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 membership inference attack that combines conventional output-side signals with aggregated routing features and applies a membership classifier learned from independently fine-tuned shadow models to the target model. Across three MoE architectures and three data domains, router telemetry consistently improves membership inference over a strong output-signal ensemble, increasing TPR at 1\% FPR by 2.7--9.4 percentage points across all nine settings. The leakage persists across full fine-tuning, frozen-router training, LoRA, and instruction tuning, and remains observable with only discrete expert selections, restricted telemetry, or a single shadow model. Mechanistic analysis further shows that the leakage does not require router-specific memorization: fine-tuning introduces membership information into hidden representations, while the router exposes a projection of this signal even when its parameters are frozen. Perturbing the telemetry reduces this additional leakage only as its fidelity degrades. Our results show that router telemetry can turn an operational signal into an additional privacy surface for fine-tuned MoE mod

Router telemetry leaks privacy information

This is a well-evidenced academic finding identifying a new privacy vulnerability in deployed MoE models, relevant to operators logging routing telemetry. The implication is specific and measured but narrow in scope, affecting model operators and fine-tuning data privacy rather than broad societal outcomes.

Router telemetry from Mixture-of-Experts (MoE) models can be used to infer whether specific data examples were used in fine-tuning, even with limited telemetry or frozen router parameters. This vulnerability significantly enhances membership inference attacks, compromising data privacy for individuals whose data was used in training.

The study demonstrates a significant privacy risk in MoE models. Router telemetry exposes membership information, increasing inference attacks. This risk is consistent across various training methods and telemetry restrictions.

SOURCE     arXiv cs.LG (Machine Learning) (tier 1)
   ↓
DOCUMENT   f4667558-e8ae-4660-a1a5-bdb55abfecd0
           https://arxiv.org/abs/2610.10616
   ↓
EVIDENCE   2 extracted excerpts
   ↓
MODEL RUN  4 runs, methodology 1.0.1
   ↓
SCORE      -82  (Strongly adverse)
   ↓
CONFIDENCE 85%