DFedReweighting: A Unified Framework for Objective-Oriented Reweighting in Decentralized Federated Learning

Kaichuang Zhang, Wei Yin, Jinghao Yang, Ping Xu

Abstract

Decentralized federated learning (DFL) has recently emerged as a promising paradigm that enables multiple clients to collaboratively train machine learning model through iterative rounds of local training, communication, and aggregation without relying on a central server which introduces potential vulnerabilities in conventional Federated Learning. Nevertheless, DFL systems continue to face a range of challenges, including fairness, robustness, etc. To address these challenges, we propose DFedReweighting, a unified aggregation framework designed to achieve diverse objectives in DFL systems via a objective-oriented reweighting aggregation at the final step of each learning round. Specifically, the framework first computes preliminary weights based on target performance metric obtained from auxiliary dataset constructed using local data. These weights are then refined using customized reweighting strategy, resulting in the final aggregation weights. Our results from the theoretical analysis demonstrate that the appropriate combination of the target performance metric and the customized reweighting strategy ensures linear convergence. Experimental results consistently show that our proposed framework significantly improves fairness and robustness against Byzantine attacks in diverse scenarios. Provided that appropriate target performance metrics and customized reweighting strategy are selected, our framework can achieve a wide range of desired learning objectives.

Introduction

Decentralized Federated Learning (DFL) represents a shift away from centralized server architectures, allowing clients—such as mobile devices or edge sensors—to train models locally and exchange updates directly with peers. While this decentralized approach mitigates "single point of failure" risks and enhances privacy, it introduces significant hurdles. Chief among these are fairness, where non-IID data distributions can lead to models that perform poorly for specific clients, and robustness, where the lack of a central authority makes the system vulnerable to malicious "Byzantine" updates.

Existing research often addresses these issues in isolation. DFedReweighting proposes a unified mechanism that handles both fairness and robustness through a single, mathematically grounded aggregation framework.

The DFedReweighting Framework

The core innovation of DFedReweighting is the relocation of system "intelligence" to the final aggregation step of each learning round. Instead of uniform or simple averaging, each client acts as a personalized filter for incoming neighbor updates.

The three-stage workflow of DFedReweighting within a decentralized federated learning system.

The framework operates in four distinct stages at each client:

  1. Auxiliary Dataset Construction ($A_k$): A small fraction of local data is randomly sampled to serve as a validation set.
  2. Preliminary Weight Calculation ($M_k$): Neighbor updates are evaluated against this auxiliary set using a Target Performance Metric (TPM), such as local accuracy or loss.
  3. Final Weight Refinement ($V_k$): A Customized Reweighting Strategy (CRS) is applied to the TPM scores. This allows the client to prioritize updates that align with specific goals (e.g., fairness or security).
  4. Aggregation: The client produces its new local model by computing a weighted average of the received updates.

Objective-Oriented Aggregation

By decoupling the measurement (TPM) from the weighting logic (CRS), DFedReweighting functions as a versatile tool for various DFL objectives.

Conceptual overview of objective-oriented aggregation for fairness and robustness.

For fairness, a client might use accuracy as the TPM and a temperature-scaled softmax as the CRS to emphasize neighbor models that perform well on its local distribution. For robustness, a client might use loss as the TPM and a clipping function as the CRS to zero out updates with suspiciously high loss, effectively neutralizing Byzantine attacks.

Experimental Evaluation

The framework was rigorously tested on standard computer vision benchmarks, MNIST and Fashion-MNIST. To simulate realistic DFL environments, the researchers employed LabelSkew and Dirichlet distributions ($Diri(\alpha)$) to create non-uniform data partitions. Robustness was tested against three Byzantine attack vectors: Gaussian Attack (GA), Sign-Flipping (SF), and A Little Is Enough (ALIE).

Key Results

The experimental results highlight the framework's dual-purpose efficacy:

  • Enhanced Fairness: In highly non-IID scenarios, DFedReweighting significantly reduced accuracy variance across clients. For instance, on MNIST ($Diri(0.1)$), variance dropped from 380.6 in standard DFL to 8.7, ensuring equitable performance across the network.
  • Byzantine Resilience: The framework demonstrated superior defense against malicious updates. In the "Sign-Flipping" attack on Fashion-MNIST, standard DFL accuracy fell to ~17%, while DFedReweighting maintained 82% accuracy, outperforming traditional robust aggregators like Krum and Trimmed Mean.
  • Efficiency via Sampling: The authors demonstrated that using a small sample for the auxiliary dataset is sufficient for effective vetting, minimizing the computational overhead on edge devices.

Conclusion

DFedReweighting provides a robust, unified solution for the inherent challenges of decentralized federated learning. By empowering individual clients to evaluate and weight neighbor updates based on local ground truth, the framework achieves both fairness and security. Future work will explore extending this framework to other objectives, such as power-aware learning and communication efficiency.

Summary updated

Abstract

Decentralized federated learning (DFL) has recently emerged as a promising paradigm that enables multiple clients to collaboratively train machine learning model through iterative rounds of local training, communication, and aggregation without relying on a central server which introduces potential vulnerabilities in conventional Federated Learning. Nevertheless, DFL systems continue to face a range of challenges, including fairness, robustness, etc. To address these challenges, we propose \textbf{DFedReweighting}, a unified aggregation framework designed to achieve diverse objectives in DFL systems via a objective-oriented reweighting aggregation at the final step of each learning round. Specifically, the framework first computes preliminary weights based on \textit{target performance metric} obtained from auxiliary dataset constructed using local data. These weights are then refined using \textit{customized reweighting strategy}, resulting in the final aggregation weights. Our results from the theoretical analysis demonstrate that the appropriate combination of the target performance metric and the customized reweighting strategy ensures linear convergence. Experimental results consistently show that our proposed framework significantly improves fairness and robustness against Byzantine attacks in diverse scenarios. Provided that appropriate target performance metrics and customized reweighting strategy are selected, our framework can achieve a wide range of desired learning objectives.
BibTeX
@misc{zhang2025dfedreweighting,
  author = {Kaichuang Zhang and Wei Yin and Jinghao Yang and Ping Xu},
  title = {{DFedReweighting: A Unified Framework for Objective-Oriented Reweighting in Decentralized Federated Learning}},
  year = {2025},
  eprint = {2512.12022},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.12022},
}
Figure