FedPCD: Priority-Constrained Federated Aggregation

Yousef Irshaid; Dara Varam; Mohamed I. AlHajri

cs.LGOct 9, 2026

FedPCD: Priority-Constrained Federated Aggregation

Authors: Yousef Irshaid, Dara Varam, Mohamed I. AlHajri

Organizations: Department of Computer Science & Engineering, American University of Sharjah · Senseable City Laboratory, Massachusetts Institute of Technology

Abstract

Federated learning (FL) seeks to optimize a shared global model from decentralized client data. While effective for average performance, this process can produce unfair performance across clients, particularly under statistically heterogeneous data distributions. Existing fairness-aware and multi-objective FL methods address this through client re-weighting, gradient correction, or common-descent formulations, but often change the global optimization direction or treat client objectives symmetrically. We propose FedPCD, a priority-constrained federated optimization framework that separates the global optimization objective from client protection. FedPCD preserves the conventional federated aggregate as the primary objective, while allowing only the minimum correction necessary to meet client performance constraints. Across heterogeneous FL benchmarks, FedPCD consistently outperforms state-of-the-art baselines, achieving the highest global accuracy across all configurations and leading worst-10% client accuracy in nearly all. FedPCD further preserves competitive trade-offs between global accuracy and client performance variability. Beyond this trade-off, FedPCD demonstrates faster empirical convergence, reaching matching or superior performance in substantially fewer communication rounds.

Figures & tables

Panel of Figure 1 — Figure 1: A taxonomy of the fairness-aware federated learning methods discussed in Sec. 2 .

Figure 2: FedPCD system diagram, showing an example with three participating clients at communication round $t$ . Left: Algorithmic steps taken to obtain a final direction in gradient space ( $\tilde{d}^{\star}$ ). Right: Objective space view, showing the resultant $\tilde{d}^{\star}$ . The half-space projection of each gradient is used to find the feasible region, which then allows us to see the minimum distortion required to satisfy all secondary constraints.

Figure 3: Demonstration of history for FedPCD, with $N=30$ clients, $10\%$ participation and $H=10$ . Top: PCD program at 4 selected rounds, marked on the curve. Paler arrows are older cache entries, and red arrows are minority clients that do not satisfy the PCD condition. Bottom: number of secondary constraints per round, which grows from 3 participants as a function of the communication rounds.

CIFAR-10
Method	5% participation									10% participation
	$\alpha=0.01$			$\alpha=0.1$			$\alpha=1$			$\alpha=0.01$			$\alpha=0.1$			$\alpha=1$
	$\bar{a}\!\uparrow$	$\sigma_{c}\!\downarrow$	$\mathrm{W}_{10\%}\!\uparrow$	$\bar{a}\!\uparrow$	$\sigma_{c}\!\downarrow$	$\mathrm{W}_{10\%}\!\uparrow$	$\bar{a}\!\uparrow$	$\sigma_{c}\!\downarrow$	$\mathrm{W}_{10\%}\!\uparrow$	$\bar{a}\!\uparrow$	$\sigma_{c}\!\downarrow$	$\mathrm{W}_{10\%}\!\uparrow$	$\bar{a}\!\uparrow$	$\sigma_{c}\!\downarrow$	$\mathrm{W}_{10\%}\!\uparrow$	$\bar{a}\!\uparrow$	$\sigma_{c}\!\downarrow$	$\mathrm{W}_{10\%}\!\uparrow$
FedAvg	48.09	19.55	18.74	60.96	13.68	37.08	71.93	5.78	61.68	57.47	15.93	29.93	64.43	11.67	43.27	72.94	5.65	63.28
AFL	26.82	18.50	7.19	46.20	10.83	25.93	64.08	5.64	54.42	43.86	11.37	22.16	53.79	8.97	36.99	65.49	5.53	56.18
q-FFL	50.62	17.16	24.00	60.30	13.12	37.95	71.18	5.75	61.03	56.60	15.02	30.21	63.31	11.43	42.77	71.96	5.61	62.04

Table 1: Results on CIFAR-10, CIFAR-100, and Fashion-MNIST.

\bar{a}

: global accuracy;

\sigma_{c}

: within-run std. across client accuracies;

\mathrm{W}_{10\%}

: worst-10% client accuracy. Within each dataset block, bold marks the best overall value in a column and underline marks the best value from the competing method group, such that the strongest FedPCD configuration and strongest baseline are both highlighted.

Figure 4: Global accuracy against client-accuracy spread at

\alpha=0.01

for both participation rates. Higher accuracy and lower

\sigma_{c}

are preferred. Crosses indicate variation across the 5 seeds

Figure 5: Client test-accuracy distributions at $\alpha=0.01$ across datasets and participation rates. Methods are ordered by global accuracy with each violin representing the kernel density of client accuracies pooled over the 5 seeds.

Figure 6: Client test accuracy distributions for FedPCD, FedAvg, and the strongest non-FedPCD baseline at

\alpha=0.01

. The distribution means are shown as dashed lines and their values are reported per panel.

Figure 7: Global accuracy against communication rounds at

\alpha=0.01

. Dots mark when FedPCD reaches FedAvg’s final accuracy, while stars mark when it reaches the strongest baseline’s final accuracy. When the latter is not reached, a hollow ring marks when that baseline reaches FedAvg’s final accuracy.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Parameters
AFL	$\eta_{\lambda}\in\{0.01,\,0.1,\,0.5\}$
q-FFL	$q\in\{0.1,\,1,\,10\}$
Ditto	$\lambda\in\{0.1,\,1,\,2\}$
FedMGDA+	$\epsilon\in\{0.1,\,0.5,\,1\}$
FedFV	$H=10,\quad\alpha\in\{0.1,\,0.3,\,0.5\}$
FedLF	$H=10$

Appendix

Table 2: Hyperparameter search spaces.

Figure 8: Comparison of FedPCD to its closest fair FL neighbors used as baselines in this study.

Figure 9: Client test-accuracy distributions at $\alpha=0.1$ across datasets and participation rates. Methods are ordered by global accuracy with each violin representing the kernel density of client accuracies pooled over the 5 seeds.

Figure 10: Client test-accuracy distributions at $\alpha=1$ across datasets and participation rates. Methods are ordered by global accuracy with each violin representing the kernel density of client accuracies pooled over the 5 seeds.

Figure 11: Client test accuracy distributions for FedPCD, FedAvg, and the strongest non-FedPCD baseline at

\alpha=0.1

. The distribution means are shown as dashed lines and their values are reported per panel.

Figure 12: Client test accuracy distributions for FedPCD, FedAvg, and the strongest non-FedPCD baseline at

\alpha=1

. The distribution means are shown as dashed lines and their values are reported per panel.

Figure 13: Global accuracy against communication rounds at

\alpha=0.1

Figure 14: Global accuracy against communication rounds at

\alpha=1

Figure 15: Global accuracy against client-accuracy spread at

\alpha=0.1

for both participation rates. Higher accuracy and lower

\sigma_{c}

are preferred. Crosses indicate variation across the 5 seeds.

Figure 16: Global accuracy against client-accuracy spread at

\alpha=1

for both participation rates. Higher accuracy and lower

\sigma_{c}

are preferred. Crosses indicate variation across the 5 seeds.

CIFAR-10
Method	5% participation			10% participation
	$\alpha=0.01$	$\alpha=0.1$	$\alpha=1$	$\alpha=0.01$	$\alpha=0.1$	$\alpha=1$
FedAvg	0.8574	0.9516	0.9935	0.9256	0.9674	0.9940
AFL	0.6285	0.9476	0.9923	0.9346	0.9727	0.9929
q-FFL	0.8958	0.9545	0.9935	0.9320	0.9678	0.9940
Ditto	0.8593	0.9514	0.9933	0.9258	0.9676	0.9941

Appendix

Table 3: Jain’s fairness index over the per-client accuracies, under the same settings as Table 1 . Higher is better. Within each dataset block, bold marks the best overall value in a column and underline marks the best value from the competing method group, such that the strongest FedPCD configuration and strongest baseline are both highlighted.

Method	$\alpha=0.01$			$\alpha=0.1$			$\alpha=1$
	$\bar{a}\uparrow$	$\bar{\sigma}_{c}\downarrow$	$W_{10\%}\uparrow$	$\bar{a}\uparrow$	$\bar{\sigma}_{c}\downarrow$	$W_{10\%}\uparrow$	$\bar{a}\uparrow$	$\bar{\sigma}_{c}\downarrow$	$W_{10\%}\uparrow$
FedAvg	57.47	15.93	29.93	64.43	11.67	43.27	72.94	5.65	63.28
FedPCD	74.84	8.89	58.03	75.69	7.45	61.28	77.51	4.68	69.24
FedPCD w/o history	59.26	16.24	29.65	65.50	11.64	44.13	73.00	5.69	63.16
FedPCD τ=0	74.31	9.15	57.55	74.88	7.90	59.91	76.96	5.12	67.85
FedPCD w/o normalization	72.99	9.40	54.44	75.08	7.46	60.85	76.96	4.66	68.81

Appendix

Table 4: Component ablation on CIFAR-10 at 10% participation.

Figure 17: Impact of increasing history window size

H

on the global accuracy, the worst-10% across clients, and the number of active secondary constraints. This is reported across all three datasets, with

\tau=0.02

and

\alpha=0.01

Normalization	$\alpha=0.01$			$\alpha=0.1$			$\alpha=1$
	$\bar{a}\uparrow$	$\bar{\sigma}_{c}\downarrow$	$W_{10\%}\uparrow$	$\bar{a}\uparrow$	$\bar{\sigma}_{c}\downarrow$	$W_{10\%}\uparrow$	$\bar{a}\uparrow$	$\bar{\sigma}_{c}\downarrow$	$W_{10\%}\uparrow$
Instantaneous $\ell_{2}$	74.84	8.89	58.03	75.69	7.45	61.28	77.51	4.68	69.24
Scalar EMA ( $\beta=0.5$ )	74.77	8.98	57.61	75.54	7.66	60.68	77.46	4.83	68.77
Scalar EMA ( $\beta=0.8$ )	74.69	9.17	56.77	75.69	7.44	61.83	77.35	4.82	68.73
Scalar EMA ( $\beta=0.9$ )	74.81	9.07	57.30	75.51	7.68	60.66	77.29	4.78	68.90

Appendix

Table 5: Normalization ablation on CIFAR-10 at 10% participation and round 500. Results are averaged over five seeds with

H=10

. We use

\tau=(0.01,0.02,0.075)

for

\alpha=(0.01,0.1,1)

, respectively.

Explore similar work

May 28, 2026cs.LG

Fairness-Aware Federated Learning with Trajectory Shapley Value

Federated learning is an emerging distributed paradigm that addresses the challenges posed by heterogeneous, privacy-sensitive data. It enables multiple clients to train a model collaboratively by aggregating their local updates at a server. However, conventional aggregation schemes typically use fixed weights that fail to reflect unequal and time-varying client contributions, leading to biased and unstable learning. To improve fairness and stability, we propose the Trajectory Shapley Value (TSV), a contribution metric that evaluates how each client influences the optimization trajectory of the global model using a validation-based, temporally consistent utility. Building on TSV, we design FedTSV, an adaptive aggregation method that converts per-round evaluations into dynamic client weights, allowing the server to respond to heterogeneous and adversarial participation in real time. Experiments on benchmark datasets show that FedTSV accelerates convergence, improves robustness, and yields more equitable contribution assessments, thereby providing a principled foundation for fairness-aware federated optimization.

Daniel Kuznetsov, Ziqi Wang

Faculty of Mathematics, Ecole Normale Supérieure Paris-Saclay, 4 Av. des Sciences, 91190 Gif-sur-Yvette, France. · Chair for Dynamics, Control, Machine Learning and Numerics – Alexander von Humboldt Professorship, Department of Mathematics, Friedrich-Alexander-Universität Erlangen-Nürnberg, Cauerstrasse 11, 91058 Erlangen, Germany.

Federated Learning Aggregation Algorithmic Fairness Shapley Value Attribution Federated Learning Heterogeneous Federated Learning

Ask PDF

May 20, 2026cs.LG

Choose Wisely and Privately: Proactive Client Selection for Fair and Efficient Federated Learning

Federated Learning enables collaborative model training across decentralized data sources without data transfer. Averaging-based FL is limited by the presence of non-IID data, which negatively impacts convergence speed and final model accuracy. Conventional alternatives suffer from significant inefficiency. Clients with noisy or highly heterogeneous data contribute expensive gradient computations that are either discarded or heavily down-weighted before aggregation. These reactive approaches waste computational resources, require more communication rounds and result in unnecessary privacy exposure. In this paper, we propose a proactive client selection framework that aims to find an optimal federation of clients whose combined data match utility and fairness requirements before training begins. Our method relies on mutual information computed from differentially private contingency tables to quantify the relevance of cross-feature correlations in the union dataset. We introduce a Potential Federation Loss (PFL) over the set of fixed-size federations, which balances two objectives. Maximizing collective data utility while ensuring fair cross-features correlations to prevent group unfairness. Client selection is expressed as an optimal subset search problem over the PFL objective, which we solve using simulated annealing under strong differential privacy guarantees for clients' local statistics. Experimental results on four benchmarks show faster, fairer, and more accurate models trained on optimally found federations, compared to uniform sampling, even when state-of-the-art adaptive aggregation or sampling strategies are employed.

Adda Akram Bendoukha, Heber Hwang Arcolezi, Nesrine Kaaniche +1

Samovar, T´el´ecom SudParis, Institut Polytechnique de Paris, France · ´ETS Montr´eal · CEA-LIST

Client Selection in Federated Learning Non-IID Federated Learning Federated Learning

Ask PDF

May 16, 2025cs.LG

Heterogeneity-Aware Client Sampling for Optimal and Efficient Federated Learning

Federated learning (FL) commonly involves clients with diverse communication and computational capabilities. Such heterogeneity can significantly distort the optimization dynamics and lead to objective inconsistency, where the global model converges to an incorrect stationary point potentially far from the pursued optimum. Despite its critical impact, the joint effect of communication and computation heterogeneity has remained largely unexplored, due to the intrinsic complexity of their interaction. In this paper, we reveal the fundamentally distinct mechanisms through which heterogeneous communication and computation drive inconsistency in FL. To the best of our knowledge, this is the first unified theoretical analysis of general heterogeneous FL, offering a principled understanding of how these two forms of heterogeneity jointly distort the optimization trajectory under arbitrary choices of local solvers. Motivated by these insights, we propose Federated Heterogeneity-Aware Client Sampling, FedACS, a universal method to eliminate all types of objective inconsistency. We theoretically prove that FedACS converges to the correct optimum at a rate of

O(1/\sqrt{R})

, even in dynamic heterogeneous environments. Extensive experiments across multiple datasets show that FedACS outperforms state-of-the-art and category-specific baselines by 4.3%-36%, while reducing communication costs by 22%-89% and computation loads by 14%-105%, respectively.

Shudi Weng, Chao Ren, Ming Xiao +1

Division of Information Science and Engineering, School of Electrical Engineering and Computer Science (EECS), KTH Royal Institute of Technology, 11428 Stockholm, Sweden

Federated Learning Heterogeneous Federated Learning Client Selection in Federated Learning

Ask PDF