When Routing Reveals Membership: Privacy Leakage from MoE Router Telemetry

Not provided in the abstract

privacyMixture-of-Expertsmembership inferencetelemetry

2610.10616

Builder Relevance

80%

2h ago

Abstract

This paper investigates the privacy implications of routing information in Mixture-of-Experts language models, revealing potential membership leakage.

Reality Card

Core Claim

Router telemetry can significantly enhance membership inference attacks on fine-tuned MoE models, increasing true positive rates by 2.7–9.4 percentage points.

Method / Result

Router telemetry improves membership inference over a strong output-signal ensemble by 2.7–9.4 percentage points at 1% false positive rate.

Limitations

The study's findings may be limited by the specific architectures and data domains tested, which could affect generalizability.

Paper to code

Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.

No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.

← Back to all papers