Continual Learning with Participation Privacy: An Auditable Buffer-Aggregate Scheme

For federated and streaming learning that continuously releases intermediate models, the authors propose an auditable buffer-aggregate scheme that provides trajectory-level (ε,δ)-DP guarantees for single-edit neighboring user streams. A randomized buffer wrapper outputs bins of size between [U,2U], reducing a single-edit stream to a Hamming-style per-bin update stream, where U is calibrated by the privacy parameters (ε,δ) and explicitly links privacy and latency. A certification theorem gives when a non-adaptive Hamming-neighbor DP proof can be extended to adaptive inputs, so that standard primitives such as tree prefix sums apply.