DP-FTRL: correlated-noise private training without sampling amplification

What changed

Applied tree aggregation to clipped cumulative gradients in FTRL, using correlated noise across updates to obtain private online/deep learning without requiring privacy amplification by sampling or shuffling.

Scientific scope

Applied tree aggregation to clipped cumulative gradients in FTRL, using correlated noise across updates to obtain private online/deep learning without requiring privacy amplification by sampling or shuffling. Inherits FTRL, continual-release trees and prior private online learning. Privacy does not require convex losses, but regret/risk utility theorems do; momentum variant has empirical benefits without a formal regret guarantee in this version. Repeated participation/multiple passes require additional accounting, not unrestricted data reuse. Pinned v1 displays a 2025 cover date despite a 2021 arXiv release stamp; date anomaly recorded, not treated as later invention.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits FTRL, continual-release trees and prior private online learning. Privacy does not require convex losses, but regret/risk utility theorems do; momentum variant has empirical benefits without a formal regret guarantee in this version. Repeated participation/multiple passes require additional accounting, not unrestricted data reuse. Pinned v1 displays a 2025 cover date despite a 2021 arXiv release stamp; date anomaly recorded, not treated as later invention.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Practical and Private (Deep) Learning without Sampling or Shuffling

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction