CovFedAvg selects federated models using local low-loss curves
Original title: CovFedAvg: Local Low-Loss Curves and Federated Model Selection
On width-16 gas-sensor clients, balanced accuracy rose from FedAvg's 88.88% to 89.84%, below matched fine-tuning.
CovFedAvg keeps local models and curve controls on clients, fits a shared endpoint from aggregated gradients, and picks a candidate from distributed training losses without centralized calibration data or client-to-client samples. Exact selection is guaranteed not to raise evaluated training loss versus the previous model and the current-round FedAvg candidate, but it is not a global-optimality or differential-privacy guarantee. Across synthetic, activity-recognition, and gas-sensor settings, the authors report modest gains over FedAvg and say results do not show a curve-specific advantage over equally costly optimization.
Unconfirmed: Gains are author-reported and do not establish an advantage unique to the curve method.
- Five methods, five seeds, two widths, and 20 rounds were evaluated without calibration data.
- Matched distributed fine-tuning reached 91.65% participating-Gas macro balanced accuracy.
- Authors say end-to-end privacy needs extra mechanisms beyond the reported data flow.
From the original article
1 min read at arxiv.orgClients observing different parts of a prediction task may learn locally equivalent models with incompatible behavior outside their observed support. We study whether redundant models permit local low-loss paths toward a compatible shared model.
By Federico Lucchetti · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.