Fairness-Aware Federated Learning with Trajectory Shapley Value

Problem
This paper addresses the limitations of conventional aggregation methods in federated learning (FL), which often utilize fixed weights for client contributions. This approach does not account for the heterogeneous and time-varying nature of client data and participation, leading to biased model updates and instability in learning outcomes. The authors propose a novel metric, the Trajectory Shapley Value (TSV), to evaluate client contributions more effectively. This work is presented as a preprint and has not yet undergone peer review.

Method
The core technical contribution is the introduction of the Trajectory Shapley Value (TSV), which quantifies the influence of each client on the optimization trajectory of the global model. TSV is computed using a validation-based utility that is temporally consistent, allowing for dynamic assessment of client contributions. The authors then develop FedTSV, an adaptive aggregation method that leverages TSV to assign dynamic weights to clients based on their per-round evaluations. This method enables the server to adjust to varying client participation and data quality in real time. The paper does not disclose specific details regarding the architecture, loss functions, or the compute resources used for training.

Results
The authors conduct experiments on benchmark datasets to evaluate the performance of FedTSV against conventional aggregation methods. They report that FedTSV accelerates convergence rates, enhances robustness against adversarial client behavior, and provides fairer contribution assessments. While specific numerical results are not detailed in the summary, the authors claim significant improvements in convergence speed and stability compared to baseline methods, indicating a marked effect size in the context of federated learning.

Limitations
The authors acknowledge that their approach may still be sensitive to the choice of validation metrics used for computing TSV, which could affect the fairness and stability of the aggregation process. Additionally, they do not address potential scalability issues when applied to a large number of clients or the computational overhead introduced by the dynamic weighting mechanism. The paper also does not explore the implications of varying data distributions across clients in depth, which could further complicate the aggregation process.

Why it matters
The introduction of fairness-aware mechanisms in federated learning is crucial for ensuring equitable model training, especially in applications involving sensitive data. By providing a principled method for assessing client contributions, FedTSV lays the groundwork for more robust and fair federated optimization strategies. This work has implications for future research in federated learning, particularly in enhancing model performance in heterogeneous environments and addressing fairness concerns in distributed systems.

Authors: Daniel Kuznetsov, Ziqi Wang
Source: arXiv:2605.30336
URL: https://arxiv.org/abs/2605.30336v1