Adaptive Fair Federated Learning (AFFL)

• Adaptive Fair Federated Learning (AFFL) is a research area that integrates fairness as an adaptive training control, shifting it from post hoc diagnostics to real-time modulation.
• It employs mechanisms such as server-side weight adjustments, client selection probabilities, and adaptive learning rates to mitigate performance disparities under non-IID data.
• Practical implementations of AFFL show reduced performance variance and improved fairness metrics through techniques like dynamic multi-objective optimization and reinforcement learning–based control.

Searching arXiv for papers on Adaptive Fair Federated Learning and closely related methods. Adaptive Fair Federated Learning (AFFL) denotes a family of federated learning methods in which fairness is treated as an online control signal for training, rather than as a purely post hoc diagnostic. In this literature, “fairness” is not a single notion. It can mean more uniform performance across clients, group fairness with respect to sensitive attributes, fair client selection, equitable participation frequency, or preference-conditioned trade-offs between fairness and utility. Correspondingly, AFFL methods adapt different components of the federated pipeline: server-side aggregation weights, client-selection probabilities, local learning rates, Momentum terms, clipping rules, preference-conditioned personalized models, or synthetic-data augmentation mechanisms (Ezzeldin et al., 2021, Salazar et al., 2022, Javaherian et al., 2024).

1. Emergence and scope of AFFL

The earliest methods in this line of work concentrated on server-side fairness-aware aggregation. FairFed proposed a server-side method that remains agnostic to the applied local debiasing and changes aggregation weights round by round based on a fairness gap between local and global fairness behavior (Ezzeldin et al., 2021). FAIR-FATE extended this direction by introducing a fairness-aware aggregation update with Momentum, using a server validation set to decide which client updates are “more fair” than the current global model and a time-varying fairness coefficient $\lambda_t$ that increases over rounds (Salazar et al., 2022).

Subsequent work expanded the adaptive component beyond aggregation alone. AdaFedAdam reformulated fair federated learning as a dynamic multi-objective optimization problem based on inverse training rates and coupled it with certainty-adaptive Adam at the server (Ju et al., 2023). FedEBA+ combined entropy-based aggregation with model and gradient alignments in a bi-level optimization framework (Wang et al., 2023). AdaFed constructed a common descent direction that decreases all clients’ losses while making higher-loss clients decrease faster (Hamidi et al., 2024). FedFair$^3$ moved the adaptive intervention to utility-aware client selection, explicitly targeting fair client selection, equitable participation frequency, and a fair accuracy distribution across clients (Javaherian et al., 2024).

By 2024–2025, the AFFL umbrella had broadened further. FedAA treated aggregation as a reinforcement-learning control problem under non-IID data and Byzantine attacks (He et al., 2024). FedFDP coupled fairness with differential privacy through fairness-aware gradient clipping and adaptive loss clipping (Ling et al., 2024). PraFFL replaced a single fairness–accuracy operating point with preference-conditioned personalized models generated by client-side hypernetworks (Ye et al., 2024). AAggFF recast fairness-aware aggregation as an online convex optimization problem with sublinear regret guarantees (Hahn et al., 2024). OTA-FFL incorporated adaptive fairness-aware aggregation into over-the-air computation (Hamidi et al., 6 Jan 2025). AFed used conditional generation and conditional GANs to approximate a global sensitive-attribute-conditioned distribution for fair data augmentation (Chen et al., 6 Jan 2025). A later medical-AI framework uses “Adaptive Fair Federated Learning (AFFL)” explicitly as a systems-level research program built around adaptive knowledge messengers, fairness-aware distillation, and curriculum-guided acceleration, though its reported gains are largely projected rather than fully established experimental outcomes (Arafat et al., 5 Oct 2025).

This suggests that AFFL is best understood not as a single algorithm, but as a research area centered on adaptive fairness control under heterogeneity.

2. Fairness objectives and formal problem settings

Most AFFL papers begin from the standard federated objective

$$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$

or closely related weighted formulations such as

$$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$

The departure from ordinary FL is that fairness enters either the objective itself, the aggregation rule, or the update geometry (Salazar et al., 2022, Ling et al., 2024).

A major branch of AFFL defines fairness as uniformity of client performance. FedAA states that a model is fairer if

$$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$

and OTA-FFL adopts the same standard-deviation proxy over client test performance (He et al., 2024, Hamidi et al., 6 Jan 2025). FedFDP uses a weighted variance of client losses,

$$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$

with smaller $\Psi$ meaning better fairness (Ling et al., 2024). FeDABoost measures fairness by the variance of client performance $\mathrm{Var}(\varphi(\mathcal{M}))$, while AdaFed emphasizes a common descent direction that is descent for every client and decreases higher-loss clients faster (Arachchige et al., 3 Oct 2025, Hamidi et al., 2024).

A second branch targets group fairness. FAIR-FATE uses Statistical Parity, Equality of Opportunity, and Equalized Odds, all with ideal value $1$: $$SP = \frac{P[\hat{Y}=1 \mid S=0]}{P[\hat{Y}=1 \mid S=1]}, \qquad EO = \frac{P[\hat{Y}=1 \mid S=0, Y=1]}{P[\hat{Y}=1 \mid S=1, Y=1]}.$$ FairFed works with Equal Opportunity Difference and Statistical Parity Difference,

$^3$0

$^3$1

and distinguishes sharply between global fairness over the federated population and local fairness on each client’s own distribution (Salazar et al., 2022, Ezzeldin et al., 2021). PraFFL uses demographic parity disparity,

$^3$2

and AFed measures discrimination by $^3$3 as well (Ye et al., 2024, Chen et al., 6 Jan 2025).

FedFair$^3$4 adds a third perspective by making fairness explicitly multi-dimensional: client selection fairness, participation-frequency fairness, and fairness in the distribution of client accuracies. Its formulation starts from

$^3$5

with the fairness argument tied not only to model outcomes but also to who participates and how often (Javaherian et al., 2024).

A common misconception is that AFFL always refers to demographic fairness. The literature does not support that restriction. Some methods are explicitly demographic, some are explicitly client-performance-oriented, and some combine participation or contribution fairness with outcome fairness.

3. Core adaptive mechanisms

The central technical idea in AFFL is that fairness is enforced by changing the training dynamics during federation. The mechanisms differ substantially across methods.

FairFed performs what it describes as closed-loop fairness control. Its per-client gap is

$^3$6

and it updates an unnormalized weight accumulator via

$^3$7

Larger $^3$8 strengthens fairness influence and can reduce accuracy (Ezzeldin et al., 2021).

FAIR-FATE separates normal and fairness-oriented updates. It computes the standard data-size-weighted update $^3$9, a fairness-aware aggregation update $$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$0 over clients whose local models are more fair than the current global model, smooths $$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$1 with a decaying Momentum term

$$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$2

and combines it with the ordinary update using

$$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$3

Here $$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$4 increases over rounds and $$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$5 decays (Salazar et al., 2022).

AdaFed uses a different route. It orthogonalizes client pseudo-gradients, scales them by loss, and computes a minimum-norm vector in the convex hull of the orthogonalized gradients. The resulting server direction satisfies the property

$$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$6

so all clients descend, but higher-loss clients descend faster (Hamidi et al., 2024).

FedEBA+ derives adaptive aggregation weights from a maximum-entropy principle: $$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$7 then blends aggregated gradients with an “ideal” aligned direction to balance fairness and global performance (Wang et al., 2023). AAggFF instead treats aggregation as a sequential decision process and updates the mixing vector online using an exponentiated-gradient-style rule

$$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$8

or an ONS-style update in the cross-silo regime, depending on the setting (Hahn et al., 2024). OTA-FFL computes round-wise adaptive weights through a modified Chebyshev scalarization constrained around the FedAvg weights, with $$\min_{\theta} f(\theta)=\sum_{k \in K} \frac{n_k}{n} G_k(\theta)$$9 controlling the fairness–accuracy trade-off (Hamidi et al., 6 Jan 2025).

FedFair$$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$0 makes adaptation act on client selection rather than only aggregation. Its utility score

$$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$1

depends on local loss, data size, training time, energy consumption, computational capability, and a penalty for excessive time, with normalized selection probability

$$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$2

The paper’s algorithm then samples clients according to this priority distribution while monitoring a resource-budget constraint (Javaherian et al., 2024).

The diversity of these mechanisms is summarized below.

Method	Adaptive mechanism	Fairness target
FairFed	Round-by-round fairness-gap-based aggregation weights	Group fairness
FAIR-FATE	Fairness-aware aggregation with Momentum, $$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$3, $$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$4	Group fairness
AdaFed	Adaptive common descent direction from orthogonalized, loss-scaled gradients	Client performance fairness
FedFair$$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$5	Utility-aware client selection with penalties and resource monitoring	Selection, participation, and accuracy-distribution fairness
FedAA	DDPG-based continuous aggregation weights and distance-based client selection	Performance fairness with robustness
FedFDP	Client-specific step-size scaling and fairness-aware clipping under DP	Balanced client fairness with privacy
PraFFL	Preference-conditioned hypernetwork for personalized trade-offs	Fairness–performance preference matching
AAggFF	Online convex optimization of aggregation weights	Client-level fairness
OTA-FFL	Modified Chebyshev round-wise adaptive weighting	Performance fairness

4. Extensions: privacy, robustness, personalization, and distribution learning

FedAA shows how AFFL can be combined with robustness against malicious clients. It uses parameter-distance-based client selection inspired by FABA, constructs the RL state from selected row-sum distances, lets the server act as a DDPG agent, and chooses continuous aggregation weights over the selected top-$$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$6 clients. Its reward is validation accuracy on a fair held-out validation set, and the paper explicitly defines robustness under Byzantine attack as higher mean test accuracy across benign clients (He et al., 2024).

FedFDP couples AFFL with differential privacy. Starting from the fairness-aware objective

$$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$7

it obtains a client-specific effective step size $$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$8 and then introduces a fairness-aware clipping coefficient

$$\mathbf{w}^* \triangleq \arg \min F(\mathbf{w}), \qquad F(\mathbf{w}) \triangleq \sum_{i=1}^N p_i F_i(\mathbf{w}).$$9

It also clips and privatizes uploaded loss values using an adaptive clipping threshold derived from the previous round’s privatized mean loss, with privacy analyzed in Rényi Differential Privacy (Ling et al., 2024).

PraFFL addresses a different issue: the assumption that each client has only a single pre-defined fairness–accuracy preference. It uses a weighted Tchebycheff scalarization and a client-side hypernetwork $$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$0 to map a preference vector $$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$1 directly to the personalized parameters $$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$2. This is adaptive at inference time because a client can supply an arbitrary preference and obtain the corresponding personalized model without retraining the full FL system (Ye et al., 2024).

AFed shifts the locus of adaptation from aggregation to distribution learning and synthetic augmentation. AFed-G trains a conditional generator on the server side in latent space; AFed-GAN trains a conditional GAN on clients and aggregates the generative parameters. The classifier is then trained with a combined accuracy and fairness objective using generated latent samples and a mixup-style fairness regularizer (Chen et al., 6 Jan 2025).

OTA-FFL incorporates AFFL into a wireless multiple-access-channel setting. The adaptive weights are recomputed each communication round from current client losses, then embedded into over-the-air analog aggregation through optimal transmit scalars and a de-noising scalar chosen to make the aggregated gradient unbiased with minimum MSE (Hamidi et al., 6 Jan 2025).

Two papers broaden the conceptual perimeter of AFFL. The medical-AI framework titled “Beyond Static Knowledge Messengers” proposes AFFL as a vision combining adaptive messenger capacity, fairness-aware distillation, and curriculum-guided acceleration; its convergence and fairness statements, as well as its reported efficiency and equity gains, are explicitly framed as projected or feasibility-study outcomes rather than a completed benchmark study (Arafat et al., 5 Oct 2025). By contrast, Adaptive FedECADO is described as not a fairness paper in the usual AFFL sense; it is better regarded as AFFL-adjacent because it adapts client and server learning rates and momentum to heterogeneity without introducing a fairness objective (Agarwal et al., 5 Oct 2025).

5. Experimental practice and reported results

The evaluation practice in AFFL is heterogeneous because the fairness targets are heterogeneous. Client-performance papers commonly report average accuracy, standard deviation or variance across clients, and worst-client or worst-10\% performance. Group-fairness papers report SP, EO, EQO, EOD, SPD, or DP disparity. Some papers add wall-clock time, hypervolume over trade-off sets, Gini coefficient, or privacy budgets (Hamidi et al., 2024, Hahn et al., 2024).

FedFair$$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$3 reported a headline reduction of 18.15% less accuracy variance on IID data and 54.78% on non-IID data, together with 24.36% less wall-clock training time on average, “without decreasing the global accuracy.” On CIFAR10, the reported variance dropped to 3.25 for FedFair$$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$4 vs 3.59 for Oort under IID and to 10.32 vs 13.13 under non-IID; similar reductions were reported for FashionMNIST and MNIST (Javaherian et al., 2024).

FairFed’s main empirical evidence is in group fairness under heterogeneity. For $$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$5, it improved EOD by about 93% on Adult with only about 0.3% accuracy drop, and by about 50% on COMPAS with similarly small accuracy cost. In the ACSIncome case study it improved EOD by about 20%, and on TILES it improved EOD from about -0.199 to 0.004 (Ezzeldin et al., 2021).

FAIR-FATE reports broad fairness gains under heterogeneity: on COMPAS it surpasses baselines in 78% of cases; on Adult it consistently surpasses baselines with about 10% average fairness improvement; on Law School and Dutch Census it surpasses baselines in 78% and 88% of cases, respectively. For strong heterogeneity with $$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$6, it reports gains such as +7% SP, +11% EO, +16% EQO on COMPAS and +8% SP, +7% EO, +19% EQO on Adult (Salazar et al., 2022).

FedFDP evaluates the privacy–fairness–utility trade-off under $$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$7. The main table reports 95.13% accuracy, $$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$8 on MNIST, 85.99% accuracy, $$\mathrm{std}\{F_k(w_1)\}_{k\in[N]} < \mathrm{std}\{F_k(w_2)\}_{k\in[N]},$$9 on FashionMNIST, and 54.21% accuracy, $$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$0 on CIFAR10. It states that, when accuracy matches the best baseline, fairness improves by 30.3% on MNIST, 6.7% on FashionMNIST, and 27.8% on CIFAR10 (Ling et al., 2024).

PraFFL evaluates the quality of fairness–performance trade-off sets by the 2D hypervolume indicator. It reports best HV values of about 0.971 on SYNTHETIC, 0.772 on COMPAS, 1.100 on BANK, and 1.010 on Adult, outperforming all baselines in the provided comparison (Ye et al., 2024).

AdaFedAdam emphasizes fairness with acceleration. On Femnist, the reported averages are 84.48% accuracy, 8.62% STD, and 74.16% worst 30% for AdaFedAdam, versus 82.97%, 11.44%, and 67.65% for FedAdam, and 77.77%, 13.20%, and 60.11% for FedAvg. On Cifar10, AdaFedAdam reports 62.81% average accuracy, 8.18% STD, and 46.01% worst 30%, improving on the listed baselines (Ju et al., 2023). AdaFed reports similarly strong fairness behavior on FEMNIST, including $$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$1 on FEMNIST-original and, on FEMNIST-skewed, $$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$2, $$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$3, and $$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$4 (Hamidi et al., 2024).

FeDABoost reports 24.4% reduction in client F1 variance on MNIST, with 0.0103 variance versus 0.0137 for FedAvg. On FEMNIST it reports median F1 0.652 and variance 0.0279, versus 0.558/0.0296 for FedAvg and 0.566/0.0317 for Ditto (Arachchige et al., 3 Oct 2025). AAggFF reports better degree of client-level fairness than existing methods in both cross-device and cross-silo settings and supplies sublinear regret guarantees for both regimes (Hahn et al., 2024). OTA-FFL reports consistent reductions in client disparity, the lowest or near-lowest standard deviation in most datasets, and notable improvements in worst-10\% accuracy (Hamidi et al., 6 Jan 2025).

6. Limitations, assumptions, and unresolved issues

Several limitations recur. Many analyses rely on restrictive optimization assumptions. FedFair$$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$5 explicitly assumes convex local losses; FedFDP assumes $$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$6-smoothness, $$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$7-strong convexity, and bounded stochastic gradients; AdaFedAdam’s proofs are for convex $$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$8-smooth losses with bounded gradients; AdaFed’s theory assumes linearly independent client gradients and standard smoothness conditions (Javaherian et al., 2024, Ling et al., 2024, Ju et al., 2023, Hamidi et al., 2024). This suggests that a substantial part of AFFL theory remains idealized relative to practical deep-learning regimes.

A second limitation concerns information requirements. FAIR-FATE requires a server-side validation set with sensitive labels. FairFed assumes secure aggregation and access to global sensitive-attribute marginals computed securely. AFed relies on sensitive attributes and latent-space generative modeling. FedFDP requires privatized uploaded loss values because the server must compute the global loss needed for fairness adjustment (Salazar et al., 2022, Ezzeldin et al., 2021, Chen et al., 6 Jan 2025, Ling et al., 2024).

A third issue is that “adaptive” itself is used in different strengths. FAIR-FATE is described as only partially adaptive, because its adaptation is mainly via predefined schedules for $$\Psi(\mathbf{w}) \triangleq \sum_{i=1}^N p_i \left(F_i(\mathbf{w})-F(\mathbf{w})\right)^2,$$9 and $\Psi$0 rather than a fully learned or feedback-driven controller (Salazar et al., 2022). By contrast, FairFed, FedFair$\Psi$1, AAggFF, FedAA, and OTA-FFL adapt weights or selection decisions from current fairness or utility signals at each round (Ezzeldin et al., 2021, Javaherian et al., 2024, Hahn et al., 2024, He et al., 2024, Hamidi et al., 6 Jan 2025).

Finally, no single fairness definition dominates the field. Performance variance across clients, worst-client accuracy, group fairness metrics, participation frequency, and preference-conditioned personalization are all treated as primary targets in different papers. This suggests that AFFL is not converging toward one canonical fairness objective; instead, it is becoming a framework for embedding different fairness notions into federated optimization under non-IID data, privacy constraints, and system heterogeneity.