Organizations: State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing, China · National Engineering Research Center of Disaster Backup and Recovery, Beijing University of Posts and Telecommunications, Beijing, China · JIUTIAN Research, China Mobile, Beijing, China · School of Cyberspace Security, Beijing University of Posts and Telecommunications, Beijing, China · University of Leicester, Leicester, United Kingdom
Multimodal data collected by heterogeneous devices are used for collaborative training, where federated learning (FL) serves as a key paradigm for effective distributed modeling with data privacy preservation. However, local training suffers from the forgetting of previously learned global knowledge under cross-client data heterogeneity, which leads to significant declines in both performance and convergence speed. Most previous studies rely on global alignment strategies to retain global knowledge, which hinder local optimization and lead to inadequate supervision of missing classes. Some studies introduce proxy datasets to supplement supervision for missing classes. However, it remains a challenge to balance class-wise global consistency and local optimization objectives without proxy datasets. In this work, we propose FedADB, a Class Anchor-Driven Dual-Branch FL framework. Specifically, the server generates class anchors optimized in a differentiable input space, which are shared across clients. These class anchors serve as global references that provide supervision for missing classes during local training. A dual-branch collaborative training mechanism is designed for clients. In this mechanism, the anchor-based global branch focuses on learning with global consistency, achieving global knowledge alignment by class-anchor balanced sampling. The local calibration branch focuses on learning discriminative local features, mitigating the degradation of local representations caused by excessive global alignment. Extensive experiments across multiple medical and natural datasets demonstrate that FedADB achieves significant improvements in both accuracy and convergence speed.
Figures & tables
Figure 1. Overview of FedADB. 1 After aggregation, the server generates class anchors and broadcasts them to the clients. 2 Each client performs dual-branch collaborative training: the AGB constructs class-balanced mixed batches from class anchors and local data, while the LCB is optimized using only local data. The shared feature extractor is jointly optimized by both branches. The updated parameters are then uploaded to the server. See caption.
Method
SVHN
CINIC-10
CIFAR-10
CIFAR-100
OrganA- MNIST
Path- MNIST
Tissue- MNIST
Average
α=0.05
α=0.1
α=0.2
α=0.5
Cross-Device Scenario ( K=100,ρ=0.05 )
FedAvg ( McMahan et al., 2017 )
84.56 (±1.57)
43.55 (±1.96)
32.32 (±1.40)
48.76 (±1.14)
57.66 (±1.43)
66.76 (±0.81)
47.54 (±0.31)
26.68 (±2.78)
69.01 (±3.19)
21.84 (±5.07)
49.87 (±1.97)
FedDyn ( Acar et al., 2021 )
83.49 (±4.28)
44.69 (±1.00)
28.38 (±0.10)
42.62 (±2.66)
66.91 (±1.91)
69.38 (±0.52)
42.87 (±0.89)
66.10 (±3.61)
74.73 (±1.59)
32.11 (±0.68)
55.13 (±1.72)
FedASAM ( Caldarola et al., 2022 )
73.86 (±3.31)
46.38 (±1.44)
30.80 (±2.04)
47.19 (±2.33)
62.63 (±0.39)
69.64 (±0.28)
46.79 (±0.85)
58.31 (±1.47)
71.85 (±1.30)
28.23 (±3.51)
53.57 (±1.69)
FedNTD ( Lee et al., 2022 )
89.23 (±0.44)
41.78 (±0.75)
39.64 (±2.17)
54.03 (±1.00)
59.53 (±0.52)
70.86 (±0.40)
46.19 (±1.23)
64.61 (±1.03)
77.31 (±1.48)
15.65 (±2.93)
55.88 (±1.20)
Table 1. Comparison of test accuracy@1 (%) among the baselines and FedADB on different datasets under cross-device and cross-silo scenarios. We set α∈{0.05,0.1,0.2,0.5} for CIFAR-10 and α=0.1 for the other datasets. Bold indicates the best result.
Figure 2. Performance of FedADB and FedAvg on CIFAR-10 ( α=0.1 ) across varying FL hyperparameter settings. See caption.
Figure 3. Performance of FedADB under different hyperparameters settings of CAG See caption.
Figure 4. Test accuracy trajectories across seven datasets ( α=0.1 ). Curves are smoothed using an EWMA with factor of 0.2, and the shaded areas represent the 95% confidence intervals based on three random seeds. See caption.
Figure 5. Class-wise accuracy of the global model over rounds. A heatmap showing the evolution of class-wise test accuracy of the global model over communication rounds. Each row represents a class, each column represents a communication round, and the color intensity indicates the corresponding test accuracy.
Figure 6. Loss landscapes of converged global models projected onto the two dominant Hessian eigendirections. See caption.
Figure 7. t-SNE visualization of class anchors and real-data representations at different training stages on SVHN. Two t-SNE plots compare class anchors with real-data representations at the early and late training stages. At the later stage, anchors move closer to their corresponding class clusters, while different classes become more clearly separated.
Figure 8. PCA-based feature distributions and correlation analysis between class anchors and real-data class prototypes at different training stages on SVHN. PCA feature distributions and correlation matrices compare class anchors with real-data class prototypes at two training stages. From the early to the late stage, within-class correlations increase, off-diagonal correlations decrease, and class-wise feature alignment becomes clearer.
Method
FedAvg
FedNTD
FedLMD
FedSOL
FedADB
Accuracy
40.35 (±0.13)
44.29 (±1.54)
48.71 (±1.03)
43.08 (±0.44)
56.30 (±0.58)
Table 6. Accuracy comparison on PACS with InceptionV1
Figure 9. Visualization and PSNR of the class anchors. Examples of class anchors generated for CIFAR-10, SVHN, and PathMNIST. The anchors contain no clearly recognizable semantic details and visually resemble high-confidence noise patterns.
Federated Learning (FL) of foundation and edge models increasingly targets deployments where client data distributions drift over time, yet existing forgetting-mitigation methods assume each client's distribution is stationary. Flashback, the strongest recent FL method against cross-client (spatial) forgetting, uses monotonically accumulating per-class label counts as a knowledge proxy; this proxy becomes miscalibrated under temporal distribution shift and anchors the global model to an outdated class balance. We formalise temporal forgetting in FL with a per-phase metric isolated from protocol-level fluctuations and propose Flashback Continual Learning (FlashbackCL), a drop-in extension of Flashback with (i) temporally-decayed label counts; (ii) a device-aware replay buffer with Class-Balanced Reservoir Sampling (CBRS); and (iii) server-side active coreset curation on the public distillation set. The results show that FlashbackCL achieves 6.9% to 10.0% relative improvement relative to Flashback, on CIFAR-10 with 50 clients and three controlled temporal shift modes, while simultaneously reducing temporal forgetting by up to 68%. A 5-variant ablation identifies CBRS replay as the critical component. FlashbackCL also improves Flashback by 3.5 points on stationary CIFAR-100, suggesting that class-balanced replay regularises spatial heterogeneity as well as temporal shift.
Mubarak A. Ojewale, Adriana E. Chis, Jorge M. Cortes-Mendoza +2
Cloud Competency Centre, National College of Ireland, Dublin, Ireland
Statistical heterogeneity limits federated learning when a single global classifier cannot represent client-specific label distributions. In this work, we propose Personalized Federated Knowledge Distillation with Head Adaptation (pFedKDH), which aggregates only the shared backbone, keeps persistent client-specific heads, and uses a recalibrated global head as a teacher during local training. Across MNIST, Fashion-MNIST, CIFAR10, and CIFAR100 under class-wise Dirichlet partitions, pFedKDH obtains the best accuracy in most settings, with accuracy gaps up to 37.67% over the weakest baseline and consistently low standard deviation across repetitions. Component-wise diagnostics and convergence results support the role of persistent heads and distillation-guided local optimization under label-skewed data.
Polycarpo Souza Neto, José Mairton Barros da Silva Júnior, Charles Casimiro Cavalcante
Universidade Federal do Cear´a, Fortaleza, CE, Brazil · Uppsala University, Uppsala, Sweden
Federated Learning (FL) enables collaborative training of distributed clients while protecting privacy. To enhance generalization capability in FL, prototype-based FL is in the spotlight, since shared global prototypes offer semantic anchors for aligning client-specific local prototypes. However, existing methods update global prototypes at the prototype-level via averaging local prototypes or refining global anchors, which often leads to semantic drift across clients and subsequently yields a misaligned global signal. To alleviate this issue, we introduce hyper-prototypes, defined by a set of learnable global class-wise prototypes to preserve underlying semantic knowledge across clients. The hyper-prototypes are optimized via gradient matching to align with class-relevant characteristics distilled directly from clients' real samples, rather than prototype-level descriptors. We further propose FedHPro, a Federated Hyper-Prototype Learning framework, to leverage hyper-prototypes to promote inter-class separability via mutual-contrastive learning with client-specific margin, while encouraging intra-class uniformity through a consistency penalty. Comprehensive experiments under diverse heterogeneous scenarios confirm that 1) hyper-prototypes produce a more semantically consistent global signal, and 2) FedHPro achieves state-of-the-art performance on several benchmark datasets. Code is available at https://github.com/mala-lab/FedHPro.
Huan Wang, Jun Shen, Haoran Li +6
School of Computing and Information Technology, University of Wollongong, Wollongong, Australia · School of Computing and Information Systems, Singapore Management University, Singapore, Singapore · Monash University, Australia +3