federated learning

To address the performance degradation in federated learning caused by data heterogeneity and communication latency—relative to centralized training—this paper proposes Federated Dual Learning (FDL), a framework tailored for hybrid settings where the server holds partial data while clients possess large-scale, heterogeneous datasets. FDL introduces a novel “server-guided–client-collaborative” dual-learning mechanism: the server leverages its local data to generate guidance signals that steer client model aggregation; it further supports asynchronous updates and tolerates cross-client non-IID data distributions. Theoretically, FDL achieves a faster convergence rate under standard assumptions. Empirical evaluation across multiple benchmark tasks demonstrates that FDL consistently outperforms state-of-the-art federated methods, effectively mitigating data heterogeneity while improving both model accuracy and training efficiency.

The theoretical foundations underlying the generalization performance of multi-step gossip (MGS) in decentralized training remain unclear, and it is unknown whether MGS can fully bridge the performance gap with centralized training. Method: This paper proposes the first unified generalization error analysis framework grounded in algorithmic stability, applicable to non-convex settings without requiring bounded gradient assumptions. It systematically characterizes the impacts of learning rate, data heterogeneity, number of nodes, local sample size, and communication topology on both generalization and optimization errors of MGS. Results: We theoretically establish that MGS achieves exponential convergence of optimization error; however, even asymptotically, an inherent generalization gap persists between decentralized and centralized training—determined solely by network topology and data heterogeneity. Empirical validation on CIFAR datasets confirms the tightness and practical guidance of our theoretical bounds.

This work addresses the vulnerability of asynchronous federated learning to malicious aggregators, which can compromise model integrity and client data privacy, thereby threatening system liveness and confidentiality. To counter this, the paper proposes the first asynchronous secure federated learning framework resilient to Byzantine aggregators. The approach leverages a replicated aggregator architecture, decoupled secure aggregation, and differential privacy via Gaussian noise, effectively mitigating Byzantine attacks without requiring consensus among aggregators. Additionally, a participation-balancing strategy is introduced to dynamically harmonize privacy budgets and model bias in asynchronous settings. Experimental results demonstrate that the proposed method maintains competitive training performance while simultaneously ensuring strong privacy guarantees, system liveness, and robustness against adversarial aggregators.

This work identifies a critical oversight in federated learning research: while existing studies emphasize network topology, they neglect the fundamental distinction between centralized (CFL) and decentralized federated learning (DFL)—namely, their training protocols (decoupled aggregation vs. joint optimization). To address this, we propose the first protocol-centric taxonomy for CFL/DFL. We systematically expose a long-standing research gap concerning distributed optimization methods in DFL and establish a theoretical triadic trade-off model among privacy, robustness, and model utility. Leveraging protocol-driven analysis, distributed optimization theory, and meta-review methodology, we develop a unified analytical framework for CFL and DFL. Our framework rigorously characterizes how distributed optimization fundamentally enables DFL, thereby providing a principled foundation and design guidance for next-generation federated learning systems that are secure, scalable, and adversarially robust.

To address three key challenges in federated learning—statistical heterogeneity (non-IID data), computational asynchrony across clients, and constrained communication bandwidth—this paper proposes the first FedAvg variant supporting *simultaneous* data heterogeneity, partial-client asynchronous updates, and gradient compression. Methodologically, it introduces a unified framework modeling all three sources of heterogeneity, integrating local asynchronous scheduling, error-compensated quantized compression, and adaptive aggregation. Theoretically, we provide a rigorous convergence analysis proving that the algorithm achieves the same asymptotic convergence rate as FedAvg under broad parameter conditions. Empirically, extensive experiments on the LEAF benchmark (up to 300 clients) demonstrate that our method significantly outperforms existing quantized and asynchronous baselines in both convergence speed and communication efficiency.

This work addresses key limitations in secure aggregation for federated learning—namely, excessive communication rounds, high computational overhead from public-key operations, and poor robustness to client dropouts—by introducing a secret sharing–based distributed aggregator architecture. In this approach, a small committee of clients acts as aggregators: each participant secret-shares its local model update among committee members, who then compute partial aggregation results locally and return shares that enable the server to efficiently reconstruct the global model. By eliminating conventional local masking and homomorphic encryption, the proposed method substantially reduces both computation and communication costs. Experimental results demonstrate that, under a realistic setting with 100,000-dimensional update vectors and 100,000 5G clients, the protocol achieves a 4.6× speedup over the OPA protocol while maintaining strong privacy guarantees and system efficiency.

This work proposes FedVG, a novel federated learning framework designed to mitigate model drift and degraded generalization caused by client data heterogeneity. FedVG introduces a global public validation set and leverages gradient norms computed on this set across individual network layers to construct client-specific scores for adaptive aggregation. Unlike conventional approaches that rely on local data volume, FedVG pioneers a layer-wise evaluation mechanism based on global validation gradients, significantly enhancing model generalization under heterogeneous conditions. Experimental results demonstrate that FedVG consistently outperforms existing baselines on both natural and medical image datasets, with particularly pronounced gains in highly heterogeneous settings, and effectively boosts the performance of mainstream federated learning algorithms.

This work proposes FedPLT, a novel federated learning approach designed to address the high communication and computational overhead, strong device heterogeneity, and issues of inconsistent parameter distributions and biased global loss estimation caused by existing partial-parameter training methods. FedPLT employs a structured partial-layer training strategy that adaptively assigns each client a personalized subset of the model based on its resource capacity. By integrating resource-aware model partitioning, hierarchical parameter selection, optimal client sampling, and aggregation optimization, FedPLT achieves performance on par with or superior to FedAvg while using only 18%–29% of trainable parameters. The method significantly reduces the number of straggler clients and demonstrates superior performance in highly heterogeneous environments compared to current state-of-the-art approaches.

This work addresses the limited robustness of federated learning under extreme conditions characterized by non-independent and identically distributed (Non-IID) client data and a majority (>50%) of malicious participants. To this end, the authors propose a heuristic defense algorithm that integrates server-side learning, client update filtering, and geometric median-based aggregation. Notably, the method operates effectively even when the server possesses only a small amount of real or synthetic data whose distribution significantly diverges from that of the clients—a setting previously unaddressed in the literature. Experimental results demonstrate that the proposed approach substantially improves model accuracy under such highly adversarial scenarios, thereby confirming its strong robustness and practical efficacy.