Uncertainty-aware Bayesian federated learning for low-resource neural machine translation at the edge

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Edge translation systems must balance limited computation, data locality, and weak performance on low-resource language pairs. This study presents Multitask Bayesian Federated Learning (MT-BayesFL), a multilingual translation framework that combines a lightweight shared encoder, task-specific decoders, matrix-normal posteriors over shared attention projections, and covariance-aware server aggregation. Client-side variational inference estimates posterior statistics that are fused by positive-semidefinite mixture moment matching, while Top-K transmission and KL regularization support communication-efficient training under non-IID data. An optional client-level differential-privacy release is defined only for a trusted or secure aggregation boundary and is reported as conditional privacy accounting rather than as part of the individual-update reconstruction diagnostic. Across five English-to-X tasks, MT-BayesFL is not uniformly strongest on high-resource pairs, where AdapterFL yields higher BLEU for English-Chinese, English-French, and English-German. Its largest descriptive advantages occur on English-Russian and English-Swahili, with BLEU scores of 32.6 and 28.3, respectively, together with an ECE of 0.082 on English-Swahili, 92.1% performance retention at 30% device dropout, a 1.786 MB upload per participating client per round, and lower reconstruction fidelity in the disclosed single-sentence no-DP diagnostic (TRA = 0.12; CosSim = 0.31). Under the separately specified aggregate release, the verified privacy accountant gives epsilon = 12.53 at delta = 0.00001; no end-to-end privacy-utility claim is made for that optional path.