FedDAF: Federated Domain Adaptation Using Model Functional Distance

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

FedDAF is a novel Federated Domain Adaptation (FDA) approach designed to improve model performance for target clients by collaborating with source clients while preserving data privacy. It specifically addresses the dual challenges of domain shifts between source and target data and limited labeled data at the target client, which most existing FDA methods neglect or assume ample target data. FedDAF achieves this by using similarity-based aggregation of a global source model and the target model. This aggregation relies on calculating the model functional distance from their mean gradient fields, computed on target data. The method normalizes the angle between these mean gradient fields using a Gompertz function. Experiments on real-world datasets, including CIFAR10, PACS, VLCS, and Office Caltech10, demonstrate FedDAF's superior test accuracy compared to existing Federated Learning (FL), Personalized Federated Learning (PFL), and other FDA methods, particularly under high domain shift and data scarcity conditions.

Key takeaway

For Machine Learning Engineers developing federated learning solutions with scarce target data and domain shifts, FedDAF offers a robust aggregation strategy. You should consider implementing its model functional distance approach to fuse source and target models, ensuring relevant information transfer based on your target client's specific objective. This method demonstrably improves test accuracy over traditional FL and PFL, even with very limited target samples.

Key insights

FedDAF improves federated domain adaptation by aggregating models based on functional distance from mean gradient fields, addressing data scarcity and domain shift.

Principles

• Model functional distance quantifies similarity for aggregation.
• Target objective guides relevant source information transfer.
• Limited target data can still yield reliable similarity measures.

Method

FedDAF aggregates a global source model and target model using a similarity weight derived from their model functional distance. This distance is computed by normalizing the angle between their mean gradient fields on target data with a Gompertz function.

In practice

• Apply model functional distance for robust federated model aggregation.
• Use Gompertz function to normalize gradient field angles.
• Prioritize target-specific objectives in cross-client model fusion.

Topics

• Federated Domain Adaptation
• Model Functional Distance
• Gradient Fields
• Data Scarcity
• Domain Shift
• Model Aggregation

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• FedIDM: Achieving Fast and Stable Convergence in Byzantine Federated Learning through Iterative Distribution Matching — FedIDM uses distribution matching and condensed data to achieve fast, stable, and robust federated learning convergence.
• Source-Free Domain Adaptation with Vision-Language Prior — DIFO++ enhances Source-Free Domain Adaptation by integrating and distilling knowledge from customized vision-language models.
• How Useful is Causal Invariance for Domain Adaptation in Finite-Sample Settings? — Causal invariance improves supervised domain adaptation in finite-sample settings when target-risk margins between candidate models are sufficiently large.
• DeFed-GMM-DaDiL: A Decentralized Federated Framework for Domain Adaptation — DeFed-GMM-DaDiL enables decentralized multi-source domain adaptation using federated GMMs and Wasserstein barycenters.
• Federated Nested Learning: Collaborative Training of Self-Referential Memories for Test-Time Adaptation — FedNL enables federated models to learn adaptive memory update rules, not just static weights, for robust test-time adaptation.
• Federated Hash Projected Latent Factor Learning — FHPLF integrates binary hash learning into federated learning to achieve superior accuracy, efficiency, and privacy by using binary gradients and projected Hamming distance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.