When Routing Reveals Membership: Privacy Leakage from MoE Router Telemetry
The research investigates a router-augmented membership inference attack against Mixture-of-Experts (MoE) models. By combining standard output signals with aggregated routing features, the method successfully identifies whether a specific example was used to fine-tune the model. Findings show that router telemetry significantly improves leakage detection across various fine-tuning methods and does not rely on memorizing specific router parameters.