Continual Learning With Participation Privacy: An Auditable Buffering-Aggregation Recipe

2026-07-08

AI Summary

This work addresses the challenge of preserving trajectory-level differential privacy in continual learning under single-edit user data modifications, which violate the Hamming adjacency assumption of standard differential privacy mechanisms. The authors propose an auditable modular framework that employs a randomized buffering wrapper to transform an adaptive stream of edits into Hamming-adjacent updates confined within independent buckets. By integrating non-adaptive differential privacy primitives, this approach achieves the first $(\varepsilon, \delta)$-trajectory-level privacy guarantee for single-edit continual learning. Key contributions include a buffer-and-aggregate mechanism, a privacy amplification theorem tailored to this setting, and an explicit quantitative trade-off between privacy and latency. The framework is compatible with standard continual learning algorithms and provides a principled parameter calibration mechanism.

Abstract

Modern federated and streaming learning systems often release intermediate models, so privacy must hold for the full trajectory under adaptive interaction. Motivated by participation privacy, we study single-edit neighboring user streams, where one insertion/deletion shifts all subsequent updates and defeats standard Hamming-neighbor continual-release analyses. We give an auditable modular recipe. A randomized buffering wrapper emits bins of size $[U,2U]$, reducing single-edit streams to a Hamming-style per-bin update stream with explicit backlog/delay guarantees, where $U$ is calibrated by the privacy parameters $(\varepsilon,δ)$. We then prove a certification theorem identifying when a non-adaptive Hamming-neighbor DP proof for a continual primitive lifts to adaptive inputs: the primitive must use fresh per-round randomness and have a stable one-round privacy profile under common adaptive context. Together, these ingredients yield trajectory-level $(\varepsilon,δ)$-DP for single-edit streams using standard primitives (e.g., tree prefix sums), with an explicit privacy--latency link via $U$.

Problem

• Continual Learning
• Participation Privacy
• Single-edit Streams
• Differential Privacy
• Adaptive Interaction

Innovation

• continual learning
• participation privacy
• differential privacy
• adaptive streams
• buffering-aggregation