Federated Multi-Objective Learning with Controlled Pareto Frontiers

Jiansheng Rao, Jiayi Li, Zhizhi Gong, Soummya Kar, Haoxuan LiarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces CR-FMOL, a novel federated multi-objective learning framework that ensures fairness across all clients by controlling the Pareto frontier. By incorporating a "preference cone" constraint, CR-FMOL guarantees Pareto-optimal solutions for every client, addressing the common issue in federated learning where majority clients are prioritized over minority ones.

What Problem Does This Paper Address?

Federated Learning (FL) trains models across decentralized devices while preserving data privacy. However, standard FL aggregation methods like FedAvg tend to favor clients with more data or common data distributions, leading to under-performance for minority clients. While multi-objective optimization (MOO) offers a way to find trade-offs between different objectives (like accuracy for different clients), existing federated MOO approaches often only guarantee Pareto optimality at the task level, not for individual clients, leaving fairness to chance. This paper aims to bridge this gap by developing a federated MOO framework that explicitly enforces client-wise Pareto optimality and fairness.

Methods

The core innovation is Conically-Regularised FMOL (CR-FMOL). It builds upon existing federated multi-gradient descent averaging (FMGDA/FSMGDA) methods. The key steps are:

  1. Local Client Updates: Each client performs local model updates using FMGDA/FSMGDA.
  2. Implicit Preference Transmission: Clients send their aggregated task-loss vectors to the server. These vectors are interpreted as implicit preferences.
  3. Cone-Constrained Server Optimization: The server then solves a novel cone-constrained Pareto-MTL sub-problem. This sub-problem is centered around a "uniform preference vector" and uses a "preference cone" (defined by an angle $\mathcal{C}$) to constrain the aggregation weights. This ensures that the resulting descent direction is Pareto-stationary for all clients within their respective cones.
  4. Controlled Pareto Frontier: This cone constraint allows for controlling the shape of the Pareto frontier, ensuring that solutions are not only Pareto-optimal but also fair across clients.

The paper also provides theoretical guarantees, proving a client-level convergence rate of $\tilde{O}(1/\sqrt{T})$ with minimal communication overhead.

Datasets

The experiments were conducted on the CIFAR-10 dataset. To simulate non-IID (non-independently and identically distributed) data distributions, a shard-based sampling strategy was employed, where clients were randomly assigned shards of data sorted by class labels. This created heterogeneous data distributions across clients while maintaining balanced dataset sizes.

Key Insights and Ideas

  • Client-wise Pareto Optimality is Achievable: The paper demonstrates that it's possible to achieve Pareto optimality not just for the aggregated model but for each individual client in a federated setting. This is a significant step towards truly fair FL.
  • The "Preference Cone" as a Fairness Enforcer: The introduction of a conical feasible set, defined by an angle $\mathcal{C}$, is a clever way to regularize the Pareto frontier. It allows practitioners to explicitly encode fairness constraints and shape the trade-off space. A smaller cone angle leads to stricter fairness, prioritizing the worst-served clients more aggressively.
  • Early-Stage Trade-off for Long-Term Fairness: CR-FMOL shows a slower initial climb in average accuracy compared to FedAvg. This is attributed to the fairness constraints, which effectively "shrink" the update norm in the early stages when client gradients disagree the most. However, this initial sacrifice leads to better fairness and comparable or even superior performance in the long run.
  • Loss vs. Fairness: The paper highlights that CR-FMOL's higher training loss compared to FedAvg is a direct consequence of its fairness objective. It's willing to sacrifice a small amount of global likelihood to boost minority clients, leading to a more conservative probability calibration.

Key Results

  • Improved Client Fairness: CR-FMOL significantly improves worst-client accuracy, with experiments showing up to an X% increase (though the specific number is missing in the provided text, the claim is made). It also leads to smaller dispersion (standard deviation) in client accuracies, indicating better fairness.
  • Comparable Final Performance: While CR-FMOL is initially slower, it is expected to catch up with FedAvg around 1,500-2,000 training rounds and maintain comparable accuracy thereafter.
  • Effective Cone Angle Control: Experiments show that a smaller cone angle ($\mathcal{C}$) leads to faster convergence of fairness metrics (lower standard deviation of client accuracies).
  • Negligible Communication Overhead: The method achieves these fairness gains with minimal additional communication cost.

Limitations and Future Work

  • Early-Stage Performance Gap: The paper acknowledges that CR-FMOL's early-stage performance (loss and accuracy) is slightly inferior to FedAvg. While they expect it to catch up, this initial lag might be a concern in time-sensitive applications.
  • Computational Cost: While communication is efficient, the server-side optimization (solving the cone-constrained QP) might introduce computational overhead compared to simpler aggregation methods.
  • Hyperparameter Tuning: The paper mentions a two-stage hyperparameter search, suggesting that finding optimal parameters for CR-FMOL might require careful tuning.
  • Future Work: The authors suggest reproducing their results with larger computational budgets (more than 1,000 global rounds) to fully observe the long-term benefits of CR-FMOL, especially regarding fairness convergence. They also invite the community to explore larger datasets and more complex FL scenarios.

Summary updated

Abstract

Federated learning (FL) is a widely adopted paradigm for privacy-preserving model training, but FedAvg optimise for the majority while under-serving minority clients. Existing methods such as federated multi-objective learning (FMOL) attempts to import multi-objective optimisation (MOO) into FL. However, it merely delivers task-wise Pareto-stationary points, leaving client fairness to chance. In this paper, we introduce Conically-Regularised FMOL (CR-FMOL), the first federated MOO framework that enforces client-wise Pareto optimality through a novel preference-cone constraint. After local federated multi-gradient descent averaging (FMGDA) / federated stochastic multi-gradient descent averaging (FSMGDA) steps, each client transmits its aggregated task-loss vector as an implicit preference; the server then solves a cone-constrained Pareto-MTL sub-problem centred at the uniform vector, producing a descent direction that is Pareto-stationary for every client within its cone. Experiments on non-IID benchmarks show that CR-FMOL enhances client fairness, and although the early-stage performance is slightly inferior to FedAvg, it is expected to achieve comparable accuracy given sufficient training rounds.
BibTeX
@misc{rao2025federated,
  author = {Jiansheng Rao and Jiayi Li and Zhizhi Gong and Soummya Kar and Haoxuan Li},
  title = {{Federated Multi-Objective Learning with Controlled Pareto Frontiers}},
  year = {2025},
  eprint = {2508.05424},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.05424},
  howpublished = {arXiv},
}
Figure