Federated learning algorithms optimization using the fittest aggregation, adaptation and alignment models

Abstract

Federated Learning (FL) has emerged as a privacy-preserving alternative to centralized machine learning (ML), meeting regulatory standards like GDPR and POPIA. However, FL faces persistent challenges, including communication overhead, resource constraints, security vulnerabilities, and data heterogeneity. This study addresses these limitations by proposing a suite of specialized optimization frameworks tailored to three distinct paradigms: Horizontal FL (HFL), Vertical FL (VFL), and Federated Transfer Learning (FTL). For HFL, we introduce FedFiTS, a framework combining fitness-based client selection with slotted aggregation, dynamic scoring, adaptive thresholds, and cohort scheduling to enhance fairness, security, and client participation. In the VFL setting, we propose FedAVL, which integrates AutoML capabilities to streamline model selection and hyperparameter tuning, reducing system setup time while enhancing model performance for clients holding vertically distributed data. For FTL, we introduce FedFusion, a framework unifying domain adaptation, frugal labeling, and self-supervised learning with diversity-aware encoders (DivEn, DivEn-mix, DivEn-c) to guide resource-constrained clients via teacher-assisted transfer. Together, these frameworks offer scalable, privacy-conscious solutions for collaborative intelligence across healthcare, computer vision, agriculture, and environmental monitoring.