Advancements in Federated Learning: Addressing Data Heterogeneity and Domain Shifts

The recent developments in the field of federated learning and its applications in various domains such as medical image segmentation, time series generation, and histopathology image analysis indicate a strong trend towards addressing the challenges of data heterogeneity, domain shifts, and privacy preservation. Innovations are particularly focused on enhancing model generalization across unseen domains, improving the quality of data generation, and ensuring robust performance in the face of spatio-temporal data shifts. Techniques such as dynamic segmentation guided by Barlow Continuity, multi-domain time series generation with domain prompts, and federated learning frameworks incorporating semantic anchors and spatio-temporal heterogeneity are at the forefront of these advancements. These methods aim to mitigate issues like client drift, catastrophic forgetting, and domain shift, thereby enabling more effective and privacy-aware machine learning models.

Noteworthy papers include:

  • A novel method for dynamic segmentation in histopathology that significantly improves dice scores by addressing both client drift and catastrophic forgetting.
  • TimeDP, a multi-domain time series diffusion model, which sets a new standard for in-domain and unseen domain generation quality.
  • FedSA, a framework that introduces semantic anchors to achieve consistent and discriminative prototypes in federated learning, outperforming existing methods.
  • STHFL, which tackles spatio-temporal heterogeneity in federated learning with a Global-Local Dynamic Prototype framework, showing remarkable results.
  • An Encoded Spatial Multi-Tier Federated Learning approach that enhances predictive accuracy across diverse datasets and spatial attributes.
  • FedSemiDG, a domain generalized federated semi-supervised learning framework for medical image segmentation, demonstrating robust generalization on unseen domains.
  • I$^2$PFL, a federated prototype learning method that mitigates domain shifts by incorporating intra- and inter-domain prototypes, showing superior performance across multiple datasets.

Sources

Federated-Continual Dynamic Segmentation of Histopathology guided by Barlow Continuity

TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts

FedSA: A Unified Representation Learning via Semantic Anchors for Prototype-based Federated Learning

STHFL: Spatio-Temporal Heterogeneous Federated Learning

Encoded Spatial Attribute in Multi-Tier Federated Learning

FedSemiDG: Domain Generalized Federated Semi-supervised Medical Image Segmentation

Mitigating Domain Shift in Federated Learning via Intra- and Inter-Domain Prototypes

Built with on top of