Multi-Source to Multi-Target Decentralized Federated Domain Adaptation

Heterogeneity across devices in federated learning (FL) typically refers tostatistical (e.g., non-i.i.d. data distributions) and resource (e.g.,communication bandwidth) dimensions. In this paper, we focus on anotherimportant dimension that has received less attention: varyingquantities/distributions of labeled and unlabeled data across devices. In orderto leverage all data, we develop a decentralized federated domain adaptationmethodology which considers the transfer of ML models from devices with highquality labeled data (called sources) to devices with low quality or unlabeleddata (called targets). Our methodology, Source-Target Determination and LinkFormation (ST-LF), optimizes both (i) classification of devices into sourcesand targets and (ii) source-target link formation, in a manner that considersthe trade-off between ML model accuracy and communication energy efficiency. Toobtain a concrete objective function, we derive a measurable generalizationerror bound that accounts for estimates of source-target hypothesis deviationsand divergences between data distributions. The resulting optimization problemis a mixed-integer signomial program, a class of NP-hard problems, for which wedevelop an algorithm based on successive convex approximations to solve ittractably. Subsequent numerical evaluations of ST-LF demonstrate that itimproves classification accuracy and energy efficiency over state-of-the-artbaselines.