Paper deep dive
Assessing the Impacts of Imperfect Datasets on Client Selections in Federated Learning
Yuan-Heng Tsai, Li-Hsing Yen, Yan-Wei Chen
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Federated learning (FL) is a popular distributed learning framework where multiple clients perform local training and a server aggregates the locally updated models. FL enables decentralized training while preserving the privacy of clients' datasets. However, non-independent and identically distributed (non-IID) or noisy datasets can lead to low model accuracy or high convergence latency. Precluding these clients through client selection may mitigate the problem, but heavily biased client selections may also degrade the learning performance. In this study, we first experimentally measure the impact of non-IID data (including skews in data quantity and label distribution), noisy data, and fairness in client selection on model accuracy and convergence. We then propose a privacy-preserving scoring method to assess each client's contribution in FL, with experiments conducted to demonstrate the effectiveness of the proposed assessment.
Tags
Links
- Source: https://arxiv.org/abs/2608.02250v1
- Canonical: https://arxiv.org/abs/2608.02250v1
Trouble viewing inline? Open PDF directly â
Full Text
30,722 characters extracted from source content.
Expand or collapse full text
Assessing the Impacts of Imperfect Datasets on Client Selections in Federated Learning Yuan-Heng Tsai, Li-Hsing Yen, and Yan-Wei Chen Department of Computer Science, National Yang Ming Chiao Tung University, Hsinchu, Taiwan. AbstractâFederated learning (FL) is a popular distributed learning framework where multiple clients perform local training and a server aggregates the locally updated models. FL enables decentralized training while preserving the privacy of clientsâ datasets. However, non-independent and identically distributed (non-IID) or noisy datasets can lead to low model accuracy or high convergence latency. Precluding these clients through client selection may mitigate the problem, but heavily biased client selections may also degrade the learning performance. In this study, we first experimentally measure the impact of non-IID data (including skews in data quantity and label distribution), noisy data, and fairness in client selection on model accuracy and convergence. We then propose a privacy-preserving scoring method to assess each clientâs contribution in FL, with experi- ments conducted to demonstrate the effectiveness of the proposed assessment. I. INTRODUCTION Federated learning (FL) enables decentralized model train- ing by FL clients while ensuring the clientsâ privacy. An FL training task begins with a central server initializing a global model and sending it to all clients. Each client then uses its local data to train the model, computes an update such as gradients or model parameters, and sends back the local update to the server. The server aggregates local updates to update the global model and then redistributes the updated global model to all clients. This completes one round of FL and the process repeats until convergence. FL is challenged by several issues. Particularly, the global model may not perform well when clients have varying data distributions known as non-independent and identically distributed (non-IID) data. Non-IID data refers to significant skews in label, feature, or quantity distributions in the dataset. It can deteriorate the model accuracy and increase conver- gence latency [1], particularly when clients have severe non- IID datasets. Another challenge comes from noisy data or mislabeled samples, which can also reduce the quality of the aggregated model. Some studies mitigate statistical heterogeneity by preclud- ing or down-weighting clients with highly non-IID data to stabilize convergence [2, 3]. In contrast, recent fairness-aware approaches argue that promoting balanced or fair client par- ticipation can also enhance convergence and improve overall model accuracy [4]. Moreover, it is still an open question how to predict a clientâs contribution to FL under this circumstance without privacy leakage. Addressing this question can help design a sophisticated client selection scheme. As most prior research focuses on partial aspects of this issue, we comprehensively analyze the challenges of non- IID and noisy datasets. We also study experimentally whether fairness in client selections can significantly affect the model quality and convergence time. We then present a scoring method to assess each clientâs importance in client selections considering both dataset quality and fairness. The rest of this paper is organized as follows. Sec. I introduces the background and related work on FL. We present experiments in Sec. 3 to study the impact of non-IID datasets and unfair client selections on test accuracy and convergence. We propose an assessment method to measure each clientâs contribution without breaching privacy in Sec. 4. Sec. 5 concludes this paper. I. BACKGROUND AND RELATED WORK A. Non-IID Datasets in Federated Learning Many studies reported that non-IID data degrades the model performance and convergence time. The non-IID setting can fall into several classes: quantity skew, label skew, and feature skew [5, 2]. 1) Quantity Skew: Quantity skew refers to a skewed dis- tribution in sample quantity across clients. In FL, if data are IID across clients and the server uses FedAvg [6] to weight each clientâs update by its dataset size, then the expected update of FedAvg equals the centralized gradient update on the union of all data [6]. So the global modelâs convergence and final accuracy depend only on the total dataset size, not on how those samples are distributed across clients. Therefore, quantity skew does not pose a serious problem for client selection as long as the server is aware of the distribution of sample quantities across clients. With that information, many client selection methods tend to favor clients with larger datasets [3]. 2) Label Skew: Label skew refers to a skewed (imbalanced) distribution in label across clients, which can cause significant model weight divergence [2]. A study [7] recommends aug- menting local datasets with extra samples to balance the label distributions. Other approaches have been proposed to prevent aggregating highly divergent models due to label skew. Ex- amples are Personalized Federated Learning (PFL) [8, 9] and Clustered Federated Learning [10, 11]. Some studies [12, 13] reduced the impact of selecting clients with highly label- skewed datasets by lowering their weights in the aggregation. Many approaches have employed deep reinforcement learning to optimize client selection in the presence of label-skewed or otherwise non-IID data distributions [12, 14]. arXiv:2608.02250v1 [cs.LG] 3 Aug 2026 3) Feature Skew: Feature skew refers to the situation where clients possess data with different feature distributions. For in- stance, in the MNIST dataset [15], a digit might be handwritten in various styles, while in the CIFAR-10 dataset [16], animals could be photographed from different angles. This results in unique feature sets for the same labels. Recent studies [5, 2] show that feature-skew is a significant non-IID issue in FL, characterized by varying relationships between features and labels across different clients. Different from label skew, which can be statistically analyzed, feature skew poses a more significant challenge for quantification due to its ambiguous definitions and widespread occurrence. In light of this, we do not view feature skew negatively but as an essential element of data diversity that ensures effective model generalization across diverse feature representations within the same label. Therefore, it is crucial to implement fairness in client selection. Such fairness ensures that the model adapts to and benefits from the intrinsic variations within the data, enhancing its robustness and general appli- cability. B. Mislabeled Samples Client datasets may contain mislabeled samples that sig- nificantly degrade the performance of the global model. For deliberate poison attacks, some studies [17] used the similarity between the local update gradient and the global update gradi- ent to gauge the extent of mislabeling. For noisy labels, some studies [18, 19] detected and filtered out mislabeled samples from local datasets. Another way to mitigate mislabeling is to add additional information to the local datasets [20, 21]. If clients can be trusted, the server can measure clientsâ mislabel levels based on a pre-trained model to evaluate the local models [22, 23, 24]. Many studies [17, 22, 24] mitigated the impact of mislabeling by adjusting the aggregation weights for clients that may contain mislabels. C. Fairness in Client Selections An uneven selection can cause the global model to learn predominantly from a small subset of clients, resulting in rapid convergence to an undesired local optimum. There- fore, fairness, which ensures that all clients have enough opportunities to participate, can potentially enhance model performance [25]. This is why some studies considered both effective participation and fairness in client selections [26]. Many studies [27, 28, 29] implemented long-term fairness constraints and used Lyapunov optimization to ensure each clientâs average participation rate. I. IMPACT OF IMPERFECT DATA AND UNFAIR CLIENT SELECTIONS We conducted several experiments to study how imperfect data (including skewed distributions in label and sample quantity and datasets containing mislabeled samples) as well as unfair client selections affect the model performance and convergence speed. Table I: Parameter Setting ParameterValueDescription |C|100Total number of clients N max 10Maximum number of clients to be selected n10Number of labels I loc 10Number of local epochs B loc 64Training batch size Îł loc 0.005Local training learning rate Table I: Six Scenarios of Label Skews NameDescription IIDEach client possessed an equal distribution of all labels. (no label skew) SingleLabelEach client exclusively possessed samples from a single label. (extreme label skew) TwoLabelsEach client had samples from two labels with an equal sample count for both labels. 70%TwoLabels30% of clients followed the IID setting and the remaining 70% followed the TwoLabels setting. 50%TwoLabels50% of clients used IID settings and 50% used the TwoLabels settings. 30%TwoLabels70% of clients used IID settings and 30% used the TwoLabels setting. We tested two classic datasets: MNIST and CIFAR-10. Unless otherwise specified, there were 100 clients, each with an IID dataset with 200 samples in the following experiment settings. We used the FedAvg algorithm with parameter setting shown in Table I. In each round, 10 out of 100 clients were randomly selected for participation in the training. Each result was averaged over ten experiments. A. Impact of Dataset Sizes We tested five different dataset sizes (number of samples in a dataset). All clients possessed datasets of equal size. The re- sults confirm that larger dataset sizes correlated with improved test accuracy and faster convergence, which is consistent with the observations of previous studies and centralized learning mechanisms. Additionally, datasets with sizes below a certain threshold failed to support effective learning. B. Impact of Label Skews We explored six scenarios with varying degrees of label skews (Table I) to see how label skew affects model accuracy. The result in Fig. 1b from the CIFAR-10 dataset revealed a clear trend that, as the degree of label skews increased, both test accuracy and convergence speed worsened. By contrast, the result in Fig. 1a from the MNIST dataset shows only slight performance declines under the same conditions, except in the extreme case of SingleLabel. This revealed that different tasks had varying levels of resilience to label skew. C. Impact of Mislabeled Samples We explored three common mislabeling scenarios men- tioned in the literature. The first is random mislabeling with probability y, where every sample had a probability of y being mislabeled with a random label. Fig. 2 shows the test accuracy over training rounds under different values of y. The second 02004006008001000 Round 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Accuracy IID 70%TwoLabels 50%TwoLabels 30%TwoLabels SingelLabel TwoLabels (a) MNIST 02004006008001000 Round 0 0.1 0.2 0.3 0.4 0.5 0.6 Accuracy IID 70%TwoLabels 50%TwoLabels 30%TwoLabels SingleLabel TwoLabels (b) CIFAR-10 Figure 1: Test accuracy over training rounds under different settings of label skew 050100150200250300 Round 0 0.2 0.4 0.6 0.8 1 Accuracy Probability 0% Probability 30% Probability 50% Probability 70% (a) MNIST 050100150200250300 Round 0.1 0.2 0.3 0.4 0.5 0.6 Accuracy Probability 0% Probability 30% Probability 50% Probability 70% (b) CIFAR-10 Figure 2: Test accuracy over training rounds under different probabilities of random mislabeling scenario is sequential mislabeling with degree y, where some labels were relabeled to the subsequent labels and the degree y specifies the number of such labels. For example, when y = k, all samples with labels 1, 2,...,k were relabeled as labels 2, 3,...k + 1, respectively. Fig. 3 shows the result with some values of y. The third scenario is cyclic mislabeling with degree y, which is similar to sequential mislabeling except that all samples with label k were relabeled as label 1 when y = k. Fig. 4 shows the result with some values of y. From these results, we note a marked decline in performance correlating with increased probability or degree of mislabeling. Compared with the results associated with label skews (Fig. 1), where learning was attainable albeit at a reduced efficiency, the results here demonstrated a pronounced drop in performance. This was especially evident in sequential mislabeling with y = 7 (Fig. 3), where the learning process failed. This observation 050100150200250300 Round 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Accuracy Degree 0 Degree 2 Degree 5 Degree 7 (a) MNIST 050100150200250300 Round 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 Accuracy Degree 0 Degree 2 Degree 5 Degree 7 (b) CIFAR-10 Figure 3: Test accuracy over training rounds under different degrees of sequential mislabeling 050100150200250300 Round 0 0.2 0.4 0.6 0.8 1 Accuracy Degree 0 Degree 2 Degree 5 Degree 7 (a) MNIST 050100150200250300 Round 0 0.1 0.2 0.3 0.4 0.5 0.6 Accuracy Degree 0 Degree 2 Degree 5 Degree 7 (b) CIFAR-10 Figure 4: Test accuracy over training rounds under different degrees of cyclic mislabeling (with 90 clients) 0100 200 300 400 500 Round 0 0.2 0.4 0.6 0.8 1 Accuracy Unfair Fair (a) Two-Label Non- IID 0100200300400500 Round 0 0.2 0.4 0.6 0.8 1 Accuracy Unfair Fair (b) Four-Label Non- IID 0100 200 300 400 500 Round 0 0.2 0.4 0.6 0.8 1 Accuracy Unfair Fair (c) Seven-Label Non- IID Figure 5: Test accuracy over training rounds under fair vs. unfair client selection with label-skewed datasets (MNIST) suggests that client selections should exclude clients suffering from severe mislabeling. Among the three mislabeling scenarios, the sequential mis- labeling scenario shown in Fig. 3 consistently resulted in the most significant performance degradation, highlighting its potential as a stress test for further refining our approach. Therefore, we will focus on the sequential mislabeling scenario in our upcoming experiments. D. Impact of Unfair Client Selections We explored how fairness in client selection affects the modelâs performance and convergence speed. To understand this, we conducted experiments on datasets with skewed label distributions and mislabeled samples. For skewed label distribution, 90% clients had IID datasets while the rest had samples from only two, four, or seven labels (with an equal sample count for each label). We tested two client selection policies. One policy, labeled Fair in the figures, randomly selected clients, which resulted in nearly uniform selection frequencies across clients. The other policy, Unfair, excluded all the clients with label-skewed datasets in the random client selections. Figs. 5 and 6 show the results for MNIST and CIFAR10, respectively. We observe that fair client selections significantly improved accuracy. However, more global rounds were required to achieve convergence, and the importance of fairness increased as more labels were contained in non-IID datasets. For mislabeled samples, we applied sequential mislabeling with degrees 3 and 5, respectively, to 80% of the clients. For 0 100 200 300 400 500 Round 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Accuracy Unfair Fair (a) Two-Label Non- IID 0100 200 300 400 500 Round 0 0.1 0.2 0.3 0.4 0.5 0.6 Accuracy Unfair Fair (b) Four-Label Non- IID 0100 200 300 400 500 Round 0 0.1 0.2 0.3 0.4 0.5 0.6 Accuracy Unfair Fair (c) Seven-Label Non- IID Figure 6: Test accuracy over training rounds under fair vs. unfair client selection with label-skewed datasets (CIFAR-10) 0 100 200 300 400 500 Round 0 0.2 0.4 0.6 0.8 1 Accuracy Unfair Fair (a)Degree3 (MNIST) 0 100 200 300 400 500 Round 0 0.2 0.4 0.6 0.8 1 Accuracy Unfair Fair (b)Degree5 (MNIST) 0 100 200 300 400 500 Round 0 0.1 0.2 0.3 0.4 0.5 Accuracy Unfair Fair (c)Degree3 (CIFAR-10) 0 100 200 300 400 500 Round 0.1 0.2 0.3 0.4 0.5 Accuracy Unfair Fair (d)Degree5 (CIFAR-10) Figure 7: Test accuracy over training rounds under fair vs. unfair client selection with 80% sequential mislabeling this setting, the Unfair policy excluded clients with mislabeled samples when randomly selecting clients for participation. The results are shown in Fig. 7. The experiment results shown in Figs. 7c and 7d revealed that random client selections significantly reduced perfor- mance compared with the Unfair policy, which is different from the results with the label-skew setting. The results suggest that mislabeled samples did not contribute valuable learning information. IV. EVALUATING DATASETS FOR CLIENT SELECTION Given the importance of privacy in FL, the server is strictly prohibited from accessing any distributional information from the local datasets, including label and feature distributions. Dataset size is the only exception, as it reveals little about the content or distribution of a clientâs private data. This restriction calls for a robust evaluation mechanism for client selection that can accurately assess each clientâs contribution without breaching privacy. In the following, we introduce a method to evaluate the quality of the clientsâ dataset under privacy requirements. A. Datasize Score We define C to be the set of FL clients. Let D i be the dataset of client i. Client iâs Datasize Score is d i = 1 if all clients have equal dataset sizes. Otherwise, d i = |D i |â min jâC |D j | max jâC |D j |â min jâC |D j | ,(1) which is the normalized size of each clientâs dataset, ranging from 0 to 1. 0255075100 Client ID 1 1.2 1.4 1.6 1.8 2 Quality Score Actual Value Average Value (a) MNIST, 500 sam- ples 0255075100 Client ID 1 1.2 1.4 1.6 1.8 2 Quality Score Actual Value Average Value (b) CIFAR-10, 500 samples 0255075100 Client ID 1 1.2 1.4 1.6 1.8 2 Quality Score Actual Value Average Value (c) CIFAR-10, 1000 samples Figure 8: Quality Score under the label-skew setting B. Quality Score We use Quality Score to capture the joint impact of misla- beling and label skew. Relying on clients to self-report dataset analyses is problematic for three reasons. First, clients may lack a complete or globally consistent view of their own data distributions. Second, clients may be unwilling to share such analyses due to privacy concerns. Third, and most critically, clients may deliberately manipulate their reported analyses to increase their chances of being selected. These concerns together motivate a server-based assessment approach. As suggested in [23], we take a server-based assessment where, at the end of each training round, the server evaluates the quality of each clientâs uploaded model using the serverâs own test dataset. The test accuracy serves as an indicator of the clientâs dataset quality. Let Ì acc i denote client iâs test accuracy. The Quality Score of client iâs dataset is q i = Ì acc i â min jâC Ì acc j max jâC Ì acc j â min jâC Ì acc j .(2) This scaling ensures that the client with the lowest test accuracy scores 0, while the highest scores 1. We conducted two experiments to validate the effectiveness of the Quality Score. 1) Label-Skew Evaluation: We used 100 clients, each with a dataset of 500 samples. The first ten clients had only one label. The number of unique labels increased by one for every subsequent group of 10 clients. The local training epochs for evaluation were set to 30. The results in Figs. 8a and 8b show that Quality Score increased with the client ID, where the clients with higher IDs possessed more labels. However, there was a performance drop in the CIFAR-10 dataset for clients ranging from IDs 70 to 80. This decline was because each clientâs total dataset size was fixed at 500 samples. Consequently, the number of samples per label decreased as the number of labels increased, leading to insufficient samples per label for effective learning. To validate this assumption, we conducted a follow-up experiment with the same settings but increased the number of samples per client to 1000 (Fig. 8c). This adjustment showed an increasing trend in performance. This confirms our hypothesis that more samples per label enhance learning effectiveness and highlights the importance of dataset size. 0255075100 Client ID 1 1.2 1.4 1.6 1.8 2 Quality Score Actual Value Average Value (a) MNIST 0255075100 Client ID 1 1.2 1.4 1.6 1.8 2 Quality Score Actual Value Average Value (b) CIFAR-10 Figure 9: Quality Score under the mislabeling setting 2) Mislabel Level Evaluation: This experiment also in- cluded 100 clients, each possessing a dataset of 500 samples. To explore different levels of mislabeling, the datasets for the first ten clients contained correctly labeled data (i.e., zero mislabeling rate). The mislabeling rate increased by 10% for every subsequent group of 10 clients. The local training epochs were also set to 30. The results shown in Fig. 9 exhibit a trend that Quality Score decreased as the client ID increased, whereas the clients with higher IDs suffered from heavier mislabeling. The results of both experiments confirm the effectiveness of Quality Score in assessing the quality of a clientâs dataset. C. Fairness Score We define a clientâs Fairness Score as a non-negative value that reflects its importance when we want to fairly select clients. It is zero for every client initially, increased by ÎČ â„ 1 per round, and decreased by 1 if the client is selected in the previous round. Specifically, let f t i be client iâs Fairness Score in round t, we have f t i = ( 0,t = 0, f tâ1 i + ÎČâ a tâ1 i , otherwise, (3) where ÎČ is a constant and a tâ1 i â 0, 1 indicates whether client i was selected in round tâ 1. Consequently, a client will have a high Fairness Score if it has not been selected for a long time. Our experiments in Sec. I-D showed that when label skew is the primary source of data heterogeneity, fairness in client selection significantly improves both accuracy and conver- gence speed. However, when clients have heavily mislabeled datasets, enforcing fairness can harm performance by includ- ing low-quality clients. Ideally, the adaptive mechanism should respond differently to these two sources of heterogeneity. However, since Quality Score captures their joint impact and cannot distinguish between them, we design the mechanism to prioritize quality over fairness when Quality Score variance is high, which is the more conservative choice given that mislabeled samples provide no learning benefit. The proposed mechanism adjusts the importance of fairness dynamically. In each round t, let C t ava denote the subset of C that are available for training. We select the top L t fair clients from C t ava with the highest Fairness Scores. The value of L t fair is L t fair = max N max ,|C t ava | α 1 P iâC t ava (q i â q) 2 |C t ava | + α 2 ! ,(4) where N max is the maximum number of clients to be selected, q i is client iâs Quality Score, and q = P iâC t ava q i /|C t ava | is clientsâ mean Quality Score in C t ava . Let C t fair â C t ava be the set of selected clients. Parameter α 2 ensures a minimum proportion of available clients are included in C t fair , while α 1 controls how sensitively L t fair expands in response to the variance of Quality Scores among available clients. When C t ava has a small variance of Quality Score, indicating that neither mislabeling nor label skew is significant among available clients, the mechanism recruits few clients with high Fairness Scores in C t fair , effectively prioritizing fairness. When the variance of Quality Score is high, indicating significant data heterogeneity among clients, the mechanism expands the candidate pool so that clients with high Quality Scores can still be selected regardless of their Fairness Scores. Note that this design prioritizes quality over fairness under high heterogeneity, which is the correct behavior when mislabeling is the dominant source of variance but may be suboptimal when label skew dominates, since Sec. I-D showed that fairness benefits model performance in the label-skew setting. V. CONCLUSIONS This study comprehensively evaluated the impacts of im- perfect datasets (specifically, quantity skew, label skew, and mislabeled samples) on model accuracy and convergence speed in FL, and experimentally showed that the optimal client selection strategy differs depending on the dominant source of heterogeneity: fairness should be emphasized under label skew but de-emphasized when mislabeling dominates. To assess client contributions without breaching privacy, we proposed a scoring mechanism comprising a Datasize Score, a Quality Score derived from server-based model evaluation, and a Fairness Score, with an adaptive mechanism that dynamically balances fairness and data quality based on Quality Score variance. A limitation of the current design is that Quality Score captures the joint impact of label skew and mislabeling but cannot distinguish between them, since high variance may arise from either source. Developing separate metrics to quantify these two factors independently is an important direction for future work. ACKNOWLEDGMENT This work was supported in part by the National Science and Technology Council of Taiwan under grant numbers NSTC 114-2218-E-A49-017 and NSTC 114-2218-E-A49-018. REFERENCES [1] Y. Huang and C. Hu, âToward data heterogeneity of federated learning,â arXiv:2212.08944, Dec. 2022. [2] Y. Zhao, M. Li, L. Lai, N. Suda, D. Civin, and V. Chandra, âFederated learning with non-IID data,â arXiv:1806.00582, 2018. [3] Y. Cho, J. Wang, and G. Joshi, âTowards understanding biased client selection in federated learning,â in Proc. 25th Intâl Conf. on Artificial Intelligence and Statistics, vol. 151, Mar. 2022, p. 10 351â10 375. [4] Y. Shi, Z. Liu, Z. Shi, and H. Yu, âFairness-aware client selection for federated learning,â arXiv:2307.10738, 2023. [5] Q. Li, Y. Diao, Q. Chen, and B. He, âFederated learning on non-IID data silos: An experimental study,â in IEEE 38th Intâl Conf. on Data Engineering, May 2022. [6] H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, âCommunication-efficient learning of deep networks from decentralized data,â in Proc. 20th Intâl Conf. on Artificial Intelligence and Statistics, Apr. 2017, p. 1273â1282. [7] Y. Diao, Q. Li, and B. He, âTowards addressing label skews in one-shot federated learning,â in Proc. Intâl Conf. on Learning Representations, 2023. [8] J. Mills, J. Hu, and G. Min, âMulti-task federated learning for personalised deep neural networks in edge computing,â IEEE Trans. Parallel Distrib. Syst., vol. 33, no. 3, p. 630â641, Mar. 2022. [9] Y. Huang, L. Chu, Z. Zhou, L. Wang, J. Liu, J. Pei, and Y. Zhang, âPersonalized cross-silo federated learning on non-IID data,â Proc. AAAI Conf. on Artificial Intelli- gence, vol. 35, no. 9, p. 7865â7873, Jul. 2020. [10] M. S. Islam, S. Javaherian, F. Xu, X. Yuan, L. Chen, and N.-F. Tzeng, âFedClust: Optimizing federated learning on non-IID data through weight-driven client clustering,â in IEEE Intâl Parallel and Distributed Processing Symp. Workshops, San Francisco, CA, USA, May 2024. [11] G. Long, M. Xie, T. Shen, T. Zhou, X. Wang, J. Jiang, and C. Zhang, âMulti-center federated learning: Clients clustering for better personalization,â World Wide Web, vol. 26, p. 481â500, Feb. 2023. [12] N. H. Nguyen, P. L. Nguyen, T. D. Nguyen, T. T. Nguyen, D. L. Nguyen, T. H. Nguyen, H. H. Pham, and T. N. Truong, âFedDRL: Deep reinforcement learning- based adaptive aggregation for non-IID data in federated learning,â in Proc. ICPP, 2023. [13] T. Zhang, K.-Y. Lam, and J. Zhao, âDeep reinforcement learning based scheduling strategy for federated learning in sensor-cloud systems,â Future Gener. Comput. Syst., vol. 144, p. 219â229, Jul. 2023. [14] X. Meng, Y. Li, J. Lu, and X. Ren, âAn optimization method for non-IID federated learning based on deep reinforcement learning,â Sensors, vol. 23, no. 22, 2023. [15] Y. LeCun, C. Cortes, and C. Burges, âMNIST handwritten digit database,â ATT Labs. Available: http://yann.lecun.com/exdb/mnist, 2010. [16] A. Krizhevsky, âLearning multiple layers of features from tiny images,â University of Toronto, Tech. Rep., 2009. [17] C.Fung,C.J.Yoon,andI.Beschastnikh, âMitigating sybils in federated learning poisoning,â arXiv:1808.04866, Jul. 2020. [18] S. Yang, H. Park, J. Byun, and C. Kim, âRobust federated learning with noisy labels,â IEEE Intell. Syst., vol. 37, no. 2, p. 35â43, Mar.-Apr. 2022. [19] J. Li, G. Li, H. Cheng, Z. Liao, and Y. Yu, âFedDiv: Collaborative noise filtering for federated learning with noisy labels,â arXiv:2312.12263, Feb. 2024. [20] J. Xu, Z. Chen, T. Q. Quek, and K. F. E. Chong, âFedCorr: Multi-stage federated learning for label noise correction,â in Proc. CVPR, Jun. 2022. [21] B. Zeng, X. Yang, Y. Chen, H. Yu, and Y. Zhang, âCLC: A consensus-based label correction approach in federated learning,â ACM Trans. Intell. Syst. Technol., vol. 13, no. 5, p. 1â23, Jun. 2022. [22] M. Yang, H. Qian, X. Wang, Y. Zhou, and H. Zhu, âClient selection for federated learning with label noise,â IEEE Trans. Veh. Technol., vol. 71, no. 2, p. 2193â2197, Feb. 2022. [23] Y. Deng, F. Lyu, J. Ren, and H. Wu, âAUCTION: Automated and quality-aware client selection framework for efficient federated learning,â IEEE Trans. on Parallel and Distributed Systems, vol. 33, no. 8, p. 1996â2009, Aug. 2022. [24] V. Tsouvalas, A. Saeed, T. Ozcelebi, and N. Meratnia, âLabeling chaos to learning harmony: Federated learning with noisy labels,â ACM Trans. Intell. Syst. Technol., vol. 15, no. 2, p. 1â26, Feb. 2024. [25] Y. Shi, H. Yu, and C. Leung, âTowards fairness-aware federated learning,â IEEE Trans. Neural Netw. Learn. Syst., vol. 35, no. 9, p. 11 922â11 938, Sep. 2024. [26] T. Huang, W. Lin, L. Shen, K. Li, and A. Y. Zomaya, âStochastic client selection for federated learning with volatile clients,â IEEE Internet Things J., vol. 9, no. 20, p. 20 055â20 070, Oct. 2022. [27] T. Huang, W. Lin, W. Wu, L. He, and K. Li, âAn efficiency-boosting client selection scheme for federated learning with fairness guarantee,â IEEE Trans. Parallel Distrib. Syst., vol. 32, no. 7, p. 1552â1564, Jul. 2021. [28] C. Battiloro, P. D. Lorenzo, M. Merluzzi, and S. Bar- barossa, âLyapunov-based optimization of edge resources for energy-efficient adaptive federated learning,â IEEE Trans. Green Commun. Netw., vol. 7, no. 1, p. 265â 280, Mar. 2023. [29] H. Zhu, Y. Zhou, H. Qian, Y. Shi, X. Chen, and Y. Yang, âOnline client selection for asynchronous fed- erated learning with fairness consideration,â IEEE Trans. Wireless Commun., vol. 22, no. 4, p. 2493â2506, Apr. 2023.