Vehicles Leave After One-Shot Training: Exact Federated Editing Cuts Cumulative Time 36%–44% in Mobility Simulation
Synopsis
One-shot federated learning finishes training in a single communication round, but the resulting model is frozen once devices join or leave, leaving full retraining as the fallback. F-SOSA draws a fixed-size uniform random client subset so outputs under neighboring device sets stay bounded in total variation, then rewrites curvature-aware fusion as a projection over client-additive sufficient statistics, enabling F-SOSA-U for withdrawing a contribution and F-SOSA-I for adding one. In a SUMO urban mobility simulation, both match retraining accuracy while cutting cumulative time and communication by 36%–44% and 22%–33%.
Interpretation
After one-shot federated training completes, the model no longer adapts to changes in the participating device set; when devices join or leave, the choice is full retraining or an approximate update that departs from retraining. Earlier one-shot federated methods optimize a single fusion operation and cannot handle contribution additions or withdrawals after training; approximate editing such as fine-tuning or model calibration lets deviation accumulate across successive edits. Against three one-shot federated baselines, FedFisher, FedLPA and Dense, F-SOSA has the lowest communication cost across all datasets and heterogeneity levels on MNIST, SVHN, CIFAR-10 and FEMNIST; in a SUMO urban mobility simulation, F-SOSA-I/U match retraining accuracy while reducing cumulative time and communication by 36%–44% and 22%–33%.
Editing is cheap because most requests do not touch the sampled client subset at all, so the model and intermediate states are reused unchanged. F-SOSA first draws a fixed-size uniform random subset of clients, which bounds the total-variation distance between outputs under neighboring device sets, with the stability level set by the sampling rate. F-SOSA-U checks whether the withdrawing client is in the subset and skips editing if not; F-SOSA-I runs a Bernoulli trial at the new client's inclusion probability and keeps the subset otherwise. Each triggered edit replaces one client contribution and performs one projection solve, with coordinator cost independent of how many clients have participated historically.
The fusion step becomes a projection onto client-additive sufficient statistics, so adding or removing one client only adds or subtracts its term. Curvature-aware fusion was an iterative optimization that generally had to be rerun when a single client changed, and the previous trajectory could not be reused. Written as an orthogonal projection onto an affine constraint, the model comes from solving a linear system whose coefficients are aggregated curvature statistics; the system stays consistent even when the aggregated Fisher matrix is sparse, singular or low-rank, and different dual solutions yield the same projected model. The curvature representation can be diagonal or K-FAC, trading curvature fidelity against communication and coordinator computation.
Perspective
The demonstrated scope is classification on MNIST, SVHN, CIFAR-10 and FEMNIST plus YOLOv8s detection on BDD100K under one-shot federated learning, with editing sequences from vehicle arrivals and departures in the Luxembourg SUMO scenario during the 08:00 morning rush hour; the theory assumes the active population stays within a constant factor of its initial size, while experiments use a fixed-ratio implementation to accommodate larger population changes.
Exactness is defined as statistical indistinguishability between editing and retraining for the model and all intermediate states, and relies on a shared initialization and fixed per-client training randomness; stability trades off against utility, since a larger sampling rate improves accuracy but weakens stability and raises expected editing cost. The fixed-ratio implementation used in experiments accommodates population changes beyond the bounded range assumed in the theory, so it does not exactly match the theoretical setting.