Shuffle Model Privacy for Federated Learning: Exact Asymptotic µ-GDP via χ²-Divergence and Bundled/Unbundled Separation

Puntos clave

• Exact asymptotic µ-GDP analysis shows improvements in privacy for federated learning frameworks.
• Closed-form µ² parameters derived for both bundled and unbundled shuffling methods enhance privacy controls.
• Observational study through the lens of χ²-divergence examines local randomizer effectiveness in federated settings.
• Indicates that as iterations increase, privacy guarantees upheld in standard convex and non-convex optimization regimes.