Year: 2026
Type: article
Abstract: Federated learning keeps source records at participating clients, but heterogeneous data can make a single shared model unsuitable for every user. We study a computationally compact adaptive personalization rule that blends a federated model with a locally fine-tuned model using a coefficient selected by client-held validation loss. All candidate methods reuse the same 30-round FedAvg checkpoint and local fine-tuning trajectory, isolating model selection from additional federated optimization. Using 1,797 handwritten-digit examples divided among eight simulated clients, we evaluate two label-heterogeneity settings, five independently seeded runs, and five evaluation rules including endpoint selection. At Dirichlet concentration 0.3, AdaptiveMix achieves 96.05% mean client-macro test accuracy, compared with 91.75% for FedAvg and 95.74% for endpoint selection. At concentration 1.0, AdaptiveMix achieves 94.27% versus 94.57% for fixed mixing. Paired seed differences and client-level outcomes characterize when interpolation helps. A pairwise masking identity test checks aggregate arithmetic, while data-locality and cryptographic privacy guarantees are explicitly distinguished. The results support a low-communication baseline for exploring heterogeneous personalized federated classifiers.
DOI: 10.54097/jceim.273
Authors: Niamh Farrell, Cian Murphy, Aoife Brennan
Institutions: Trinity College Dublin, University College Dublin
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7235387507
