Rising from Ashes: Generalized Federated Learning via Dynamic Parameter Reset

Regularization-Enhanced Federated Class-Incremental Learning with Generative Replay for Smart Edge

Federated learning is an emerging paradigm that addresses privacy concerns by training on decentralized data, but it faces significant performance challenges due to factors like non-independent and identically distributed (non-IID) data and dynamic data distributions. These challenges are further amplified in federated class-incremental learning under realistic, non-monotonic class streams, where previously seen classes can intermittently reappear alongside new classes—making generative replay highly susceptible to training instability, mode collapse, and feature drift. To address these issues, we propose RAVEL-FCL, a generative replay-based framework for federated continual learning. Our approach integrates an improved generative model based on Rebooting ACGAN with multi-level feature alignment to ensure consistency, and employs Elastic Variational Continual Learning on the server to probabilistically regularize the global model and preserve past knowledge. Experimental evaluations show that our method achieves stable convergence and significantly outperforms state-of-the-art methods on standard benchmarks—for example, it improves accuracy by 5.27% over FedCIL on EMNIST-Balanced and by 1.60% and 1.27% over the strongest baselines on TinyImageNet and ImageNet-LT, respectively—while maintaining a forgetting rate below 5% across tasks. By jointly addressing feature drift and catastrophic forgetting, RAVEL-FCL provides a robust and privacy-preserving solution suitable for the adaptive and autonomous systems envisioned for Industry 5.0.

Global Federated Learning Strategies for Building Efficient Personalized Models

Federated learning (FL) is a practical framework that can train models on distributed user data while guaranteeing data privacy; however, due to heterogeneity in which each user has a different data distribution, problems frequently arise where both global and personalization performance deteriorate simultaneously. This dissertation presents methodologies for building efficient personalized models by identifying which strategies are effective in the global training stage and by showing how to preserve global knowledge while securing user-specific performance during local adaptation. First, we show that as data heterogeneity increases, the collapse of feature vectors is a more fundamental bottleneck than classifier weights, and propose a method that directly mitigates the discrepancy in representation magnitude between local and global models. Second, we analyze that a training approach that strengthens local alignment can induce forgetting of global knowledge (e.g., categories not observed locally), and propose a method that achieves both local alignment and global knowledge preservation by combining feature distillation based on the global model's feature vectors. Third, in federated personalized reward model learning with preference heterogeneity, we empirically verify the conventional belief that"increasing the number of global models yields better initialization,"and we show that when sufficient local fine-tuning is allowed, a single global initialization can instead provide stronger personalization performance. This study redefines the role of global initialization under data and preference heterogeneity and provides practical training strategies that simultaneously satisfy global knowledge preservation and personalization.

FedALA-WR: personalized federated learning with adaptive local aggregation and weighted replay

FuGuard: Client-Level Federated Unlearning via Generative Surrogates and Optimal Transport.

Federated learning (FL) is a widely adopted paradigm that enables collaborative model training while preserving data privacy. As concerns around data poisoning and the "right to be forgotten" continue to grow, federated unlearning, which is the ability to remove the influence of specific training data from a trained FL model, has become increasingly critical. However, existing unlearning methods often require expensive retraining or fail to achieve good forgetting effects, limiting their practicality in real-world FL systems. In this work, we propose FuGuard, a dual-strategy federated unlearning framework, designed for efficient and ideal client-level data removal. FuGuard combines the generative surrogate, which approximates the contribution of the target client, with optimal transport regularization that softly constrains model parameter drift during unlearning. This approach effectively removes the influence of the target client while preserving the stability and performance of the global model. To evaluate the forgetting capability, we conduct testing using backdoor attacks and member inference attacks (MIAs) for residual data influence. Empirical results on different benchmarks demonstrate that FuGuard significantly reduces the impact of the target client's data while maintaining the performance of nontarget clients, consistently outperforming state-of-the-art baselines in both forgetting effectiveness and accuracy retention. Our code is accessible at: https://anonymous.4open.science/r/FuGuard-0263.

Beyond Parameter Space: NTK-Guided Personalized Aggregation for Robust Federated Learning

Federated learning (FL) enables collaborative model training across distributed clients while keeping data local. A central challenge is determining which client updates are beneficial for aggregation with respect to each client's target domain. Existing methods typically address this problem in parameter space by comparing model parameters or gradients. However, parameter-space similarity can be a poor proxy for predictive behavior, especially under heterogeneous, non-IID data. Consequently, updates that are misaligned with a client's target domain, including those caused by heterogeneous data or malfunctioning clients, may degrade local model performance. We propose Local Inference Guided Aggregation for Heterogeneous Training Environments to Yield Enhancement Through Agreement and Regularization (LIGHTYEAR), a federated learning framework that performs update selection in function space. LIGHTYEAR uses an NTK-based agreement score to characterize predictive behavior and determine a personalized aggregation set for each client. By relating model parameters to local predictive responses, the Neural Tangent Kernel (NTK) provides a more expressive criterion for update selection than parameter-space similarity alone. Because function-space information is not available before aggregation in conventional centralized FL, LIGHTYEAR uses a peer-to-peer (P2P) topology in which clients exchange updates directly and evaluate incoming models on private validation data. Each client selects only updates that are beneficial for its own target domain and aggregates them using a regularized rule that improves stability under heterogeneity. Across five datasets and nine baseline methods, LIGHTYEAR consistently outperforms centralized FL baselines and existing P2P approaches.

Federated Learning with Consistency Optimization Algorithms under Non-IID Data

: Federated learning (FL) enables collaborative training of deep neural architectures while preserving data privacy, yet its performance often deteriorates in non-IID scenarios, which stems from client-side distribution drift and divergent local updates induced by pervasive data heterogeneity. This challenge is particularly critical for maintaining the structural consistency and generalization of neural models across diverse, distributed sources with significant distribution shifts. In this paper, we investigate how to effectively mitigate label distribution shift and feature distribution skew to enhance the global representation stability of neural architectures. We propose Federated Learning with Consistency Optimization Algorithms (FedCO), a novel optimization framework that incorporates a label-skew-aware correction loss and neural feature distribution regularization during local training. Specifically, our method aligns the internal statistics of architectural components between local and global models to suppress feature-space drift. Combined with an adaptive global aggregation mechanism guided by label entropy, this approach ensures that model updates from heterogeneous clients are consistently integrated into the global architectural parameters. Experimental results on multiple benchmarks demonstrate that FedCO significantly improves accuracy and convergence under diverse non-IID settings. For instance, on CIFAR-10 with extreme heterogeneity ( α = 0.05), FedCO achieves 73.64% accuracy, outperforming FedAvg by 7.32%; it also attains the highest accuracies on CIFAR-100 (62.63%) and TinyIma-geNet (37.82%) under the same setting. Our findings provide a robust strategy for integrating distribution-aware local training with adaptive structural aggregation, offering new insights into enhancing the reliability of distributed neural systems in real-world deployments. The code related to this algorithm is available at https://github.com/Donglin0730/ FedCO.