Technical Summary: FedDAF

Problem Statement

Federated Domain Adaptation (FDA) aims to improve model performance at a target client by collaborating with source clients while preserving data privacy. However, existing FDA methods face two critical, often unaddressed challenges simultaneously:

1. Domain Shift: Inconsistencies in data distributions between source and target clients cause local model drift, leading to suboptimal global aggregation and poor generalization on the target.
2. Data Scarcity: Many existing approaches assume the target client possesses ample unlabeled data. In reality, target clients often suffer from a severe lack of labeled samples, making standard training or personalized federated learning (PFL) ineffective.

Furthermore, current methods that attempt to address both issues often fail to prioritize the sharing of relevant information from source clients in a manner tailored specifically to the target client's objective.

Methodology: FedDAF

The authors propose FedDAF (Federated Domain Adaptation using Model Functional Distance), a novel framework designed to maximize the transfer of relevant source information to the target model while aligning with the target's specific objective, even under conditions of limited data.

Core Mechanism: Model Functional Distance

Unlike methods that rely on gradient updates or raw data distribution alignment, FedDAF aggregates the global source model ($w^S$) and the target model ($w^T$) based on their similarity, quantified via Model Functional Distance. This distance is computed as follows:

1. Mean Gradient Fields: The server broadcasts the global source model to the target client. The target client computes gradient fields for both the global source model and its own local model using the target data (even if limited).
2. Cosine Similarity & Angle: The mean gradient fields of the two models are calculated. The cosine similarity between these mean gradients is computed, and the angle $\theta$ between them is derived.
3. Normalization via Gompertz Function: The angle $\theta$ is normalized using a Gompertz function to produce a similarity weight $\alpha \in [0, 1]$. This weight represents the "distance" between the models relative to the target objective.
4. Adaptive Aggregation: The adapted target model ($w_n$) is formed by a weighted combination of the previous target model and the global source model:
   $$w_n = \alpha w^S_{n-1} + (1 - \alpha) w^T_{n-1}$$
   Here, $\alpha$ acts as the aggregation weight. If the models are functionally similar regarding the target objective, the source model contributes more; if they diverge, the target model is preserved.

Training Protocol

• Target Client: Uses the adapted model $w_n$ as an initialization for local training on its limited target data. The goal is to optimize the target objective $F_T$ while maintaining a controlled distance from the global source objective $F_S$.
• Source Clients: Perform standard local training on their respective datasets and update the global source model via simple averaging on the server.
• Iteration: This process repeats over $N$ communication rounds.

Key Contributions

The paper outlines three primary contributions:

1. Dual-Challenge Solution: FedDAF is proposed to simultaneously address domain shift and limited target data in FDA, a combination often neglected by prior work.
2. Novel Aggregation Strategy: The introduction of a similarity-based aggregation method using Model Functional Distance. This metric is derived from mean gradient fields evaluated on target data, allowing for reliable similarity measurement even with scarce samples.
3. Objective-Aligned Optimization: The method ensures the adapted target model is optimized at the optimum of the target objective, positioned at an optimized distance from the global source objective. This maximizes the transfer of relevant source information specifically aligned with the target's needs.

Experimental Results

The authors validated FedDAF through extensive experiments on real-world datasets under both controlled and real-world domain shift settings.

• Datasets:
  • Controlled Shift: CIFAR-10 with varying levels of Gaussian noise (domain shift) and label scarcity (5%, 25%, 50% of training data).
  • Real-World Shift: PACS, VLCS, and Office Caltech10 datasets, simulating diverse domain shifts (e.g., photo vs. sketch, different camera types).
• Baselines: FedDAF was compared against standard Federated Learning (FedAvg), Fine-tuned FedAvg, Personalized FL methods (FedDWA), and existing FDA methods (FedGP).
• Performance:
  • On CIFAR-10, FedDAF consistently outperformed all baselines across all levels of data scarcity and domain shift.
  • On real-world datasets (PACS, VLCS, Office Caltech10), FedDAF achieved superior test accuracy compared to existing FL, PFL, and FDA methods, particularly in scenarios with extreme data scarcity (e.g., using only 2% of target training samples).
• Parameter Analysis: The study identified that the Gompertz function parameter $\mu$ influences performance based on the degree of domain shift, with $\mu=5$ performing best under high domain shift and $\mu=1$ under lower shift.

Significance and Claims

The paper claims that FedDAF represents a significant advancement in Federated Domain Adaptation by moving beyond the assumption of ample target data. Its significance lies in:

• Robustness to Scarcity: Demonstrating that effective adaptation is possible even when the target client has very few labeled samples, a scenario where centralized training and standard PFL fail.
• Target-Centric Adaptation: Unlike previous FDA methods that may aggregate information without specific regard for the target's unique objective, FedDAF explicitly constructs the aggregation weight based on the target's functional distance. This ensures that the "maximum relevant information" is fused into the target model.
• Practical Applicability: The method addresses real-world constraints found in domains like medical imaging (e.g., breast cancer diagnosis), where data is often siloed, privacy-sensitive, and sparsely labeled.

The authors conclude that by optimizing the adapted target model at the optimum of the target objective while maintaining a controlled distance from the source global objective, FedDAF effectively bridges the gap between source and target domains in privacy-preserving, data-scarce environments.