federated conditional variational autoencoder

This work addresses the performance degradation of federated learning under data heterogeneity—including label shift, covariate shift, and concept shift—as well as in data-scarce settings. The authors propose a communication-free personalization method that dynamically conditions a single global model by embedding client-specific PCA statistics, computed locally, as continuous conditional inputs. Trained within the standard federated learning framework, the approach incurs no additional communication overhead. Extensive experiments demonstrate that the method consistently outperforms existing approaches across 97 configurations, surpassing even an oracle baseline with access to true cluster assignments by 1–6% under complex heterogeneity, while maintaining robust performance under data sparsity—where it remains the only method exhibiting consistent stability.

This study addresses the challenge of client drift in federated learning caused by data heterogeneity, which severely hinders global model convergence and training efficiency. To mitigate this issue, we propose a distributed optimization framework based on latent information sharing that aggregates cross-client knowledge by exchanging a minimal set of hidden-layer activations, thereby alleviating data heterogeneity while preserving privacy guarantees. We provide theoretical convergence analysis for the proposed method and demonstrate through extensive experiments that it significantly outperforms baseline approaches such as FedProx under a fixed communication budget. Notably, the framework substantially improves model accuracy without introducing additional communication overhead, validating the effectiveness of activation sharing mechanisms for federated optimization.

This work addresses the suboptimality of fixed or heuristic aggregation weights in federated learning caused by heterogeneous client data distributions. To overcome this limitation, the authors propose a dynamic aggregation mechanism based on Conditional Random Fields (CRFs), which introduces structured probabilistic inference into the federated aggregation process for the first time. The approach employs an energy function to jointly model each client’s unary reliability and pairwise interactions among clients, enabling dynamic optimization of the aggregation weights for the global model. Evaluated under non-IID data settings, the proposed method significantly outperforms mainstream federated learning baselines, yielding notable improvements in both convergence speed and final model accuracy.

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.

To address the challenges of dynamically adapting socially appropriate behaviors and mitigating catastrophic forgetting in multi-robot federated continual learning, this paper proposes FedRoot and FedLGR—a synergistic framework. FedRoot decouples feature and task learning to enable lightweight weight aggregation, reducing CPU and GPU overhead by 86% and 72%, respectively. FedLGR introduces a novel latent-space generative replay and pseudo-feature reenactment mechanism, alleviating forgetting without storing raw data. Evaluated in a simulated living-room environment, the framework enables distributed robots to progressively and resource-efficiently learn social norms. Experiments demonstrate consistent superiority over baselines on social behavior discrimination tasks, achieving 84% and 92% reductions in computational and memory resource consumption, respectively. The framework significantly enhances the practicality and scalability of federated continual learning for real-world social robotics applications.

This work addresses the performance degradation in federated learning caused by client data heterogeneity by proposing FedGMI, a framework that effectively balances personalization and generalization. FedGMI introduces, for the first time, a probabilistic mixture modeling paradigm, representing each client’s data distribution as a convex combination of multiple shared latent distributions. A variational autoencoder (VAE) is employed as a generative density estimator to jointly infer both the mixture components and the shared distributions. Experimental results demonstrate that FedGMI accurately identifies intrinsic data distributions, estimates mixture weights, and maintains robust performance under communication constraints, significantly enhancing personalized model effectiveness and collaborative learning efficiency.

This work addresses the challenge of deploying large foundation models on resource-constrained clients in federated learning by proposing FedSLM, a novel framework that constructs self-contained lightweight client models via SVD-based low-rank decomposition. FedSLM introduces a two-stage aggregation protocol: intra-group synchronization of low-rank adapters followed by inter-group fusion of full-rank representations. The method innovatively incorporates a structure-aligned fusion mechanism operating on nested subspace manifolds, complemented by a confidence-guided auxiliary loss and a weak-to-strong knowledge distillation strategy to enable efficient knowledge transfer. Experimental results demonstrate that FedSLM significantly outperforms existing federated approaches on both natural language and vision-language tasks, maintaining strong representation capabilities under both IID and non-IID data settings while reducing client memory consumption to approximately 50% of that required by the full model.

This study addresses unsupervised anomaly detection for on-device predictive maintenance in federated learning by systematically evaluating the performance, communication overhead, and personalization trade-offs of variational autoencoders (VAEs), generative adversarial networks (GANs), and denoising diffusion probabilistic models (DDPMs) under both full-federation and partial-federation settings—such as sharing only the decoder. The work proposes the first component-level sharing taxonomy for federated generative models, formalizing partial parameter sharing as a mechanism for model personalization. Experimental results demonstrate that, under non-IID data distributions and bandwidth-constrained conditions, DDPMs with shared decoders can outperform fully federated training. Moreover, VAEs and DDPMs exhibit consistently greater stability and robustness compared to GANs; although full-federation improves GAN stability, its overall performance remains inferior.

This work addresses the challenges in heterogeneous federated learning, where significant disparities in client resources and the reliance of existing methods on predefined architectures or hypernetworks lead to coarse personalization granularity and high computational and memory overhead. To overcome these limitations, the paper proposes a decentralized dynamic module assembly paradigm that abandons fixed hypernetworks. It introduces an attention-driven local adaptive assembly mechanism (AttenAssemble), a topology-aware inter-client module sharing architecture (SymbioArchitect), and an attention-enhanced centralized training–decentralized execution strategy (CoRe-Tune) to enable fine-grained, privacy-preserving personalized model construction. Experiments demonstrate that the proposed approach improves relative accuracy by up to 13.8%, substantially reduces performance variance across clients, and significantly lowers decision latency and peak memory consumption.

This work addresses the performance and resource imbalances in federated learning caused by non-IID data distributions, device heterogeneity, and inefficient communication. To this end, the authors propose FedKDNAS, a novel framework that uniquely integrates client-side distributed neural architecture search with server-guided knowledge distillation. Each client autonomously discovers lightweight models under accuracy and resource constraints, while contributing logits to a global distillation process. Key innovations include a hybrid supervised-distillation objective, a logit-sharing mechanism over a common reference set, and server-side strategies involving prediction smoothing and teacher model fusion. Extensive experiments across six datasets demonstrate that FedKDNAS outperforms six state-of-the-art methods, achieving up to a 15% accuracy gain, 28% reduction in client CPU usage, and a 44-fold decrease in communication overhead.