A wireless FL system operating under RIS-assisted blocked-link propagation scenarios is considered, and a joint convergence-latency optimization problem is cast as a mixed-integer nonlinear programming (MINLP) problem, and solved using a low-complexity hybrid alternating optimization framework.
Abstract
Federated learning (FL) over wireless networks suffers from significant training latency and degraded convergence due to unreliable wireless transmission, especially under blocked propagation environments. Although reconfigurable intelligent surfaces (RISs) can improve communication reliability, existing wireless FL studies rarely characterize the trade-off between learning convergence and communication delay under modulation-dependent transmission errors. In this paper, we consider a wireless FL system operating under RIS-assisted blocked-link propagation scenarios, and focus on adaptive modulation and sub-channel allocation for convergence-latency aware communication design. By characterizing the effect of symbol errors on uploaded local gradients, we derive a convergence-related upper bound that reveals the impact of symbol error rate (SER) on FL loss decay. Based on this result, we formulate a joint convergence-latency optimization problem, which is cast as a mixed-integer nonlinear programming (MINLP) problem, and solve it using a low-complexity hybrid alternating optimization framework. Extensive experiments on MNIST, CIFAR-10, and Speech Commands show that the proposed scheme consistently achieves faster convergence and higher test accuracy than existing adaptive communication schemes, especially in complex tasks and challenging wireless scenarios.
Recent Federated Learning (FL)-enabled 5G Open Radio Access Network (O-RAN) systems continue to face significant challenges associated with scalability, convergence speed, dynamic power allocation, and the adaptive optimization of model weights. Furthermore, the stochastic nature of wireless channels and increasing user density often degrade latency performance and reliable model aggregation during federated communication rounds. To address these challenges, this paper proposes a hybrid Lyapunov–DLD framework that integrates Lyapunov Drift-Plus-Penalty (LDPP) control with Dual Lagrange Decomposition (DLD) to enable self-organizing network (SON)-driven adaptive Open Radio Unit (O-RU) resource allocation, thereby ensuring timely model updates between users and the server. Specifically, the LDPP mechanism dynamically provisions O-RU computational and bandwidth resources according to real-time environmental drift, while Projected Momentum Gradient Descent (PMGD) accelerates convergence and improves penalty adaptation efficiency. Leveraging the bandwidth allocated by LDPP, the DLD framework subsequently performs channel-aware transmission power allocation based on Channel State Information (CSI) to ensure reliable model weight delivery under stochastic radio conditions. In addition, the dynamic model weight generation problem at the user side is addressed using a Proximal Policy Optimization (PPO)-based framework integrated with the Douglas–Rachford Federated Learning (FeDR) model, which mitigates the impact of non-IID data distributions arising from heterogeneous user datasets and accelerates the convergence of the local loss function through adaptive local model parameter optimization. Experimental results demonstrate that the proposed framework improves convergence, accuracy, scalability, and signal-to-noise ratio (SNR), data rate, while simultaneously reducing latency, and energy consumption for both CIFAR-10 and FEMNIST datasets compared with the FedProx, FedADMM, LyFeD (Lyapunov-based federated learning) and FL-MEC benchmark schemes.
Kofi Kwarteng Abrokwa, Qi Jiang, Zhou-Qin Ma et al.· IEEE Transactions on Green C...· 0 citations
Federated learning (FL) enables multiple devices to collaboratively train a global model without sharing local data. However, due to limited local computing capability and communication bandwidth, FL suffers from high learning latency, especially when the model size is large. To address these issues, we propose APQ-FL, an adaptive model pruning and quantization method for wireless FL, to reduce the neural network size and improve communication efficiency. Moreover, device selection and wireless resource allocation are also integrated. We first present a convergence analysis of FL with model pruning and quantized transmission, and then jointly optimize the pruning ratio, quantization bit width, device selection, and wireless bandwidth allocation to minimize the convergence upper bound under latency and bandwidth constraints. We prove that the optimized quantization bit width can be obtained via binary search, and derive the closed-form solutions for the optimal pruning ratio and bandwidth allocation. Subsequently, we propose an efficient device selection strategy and further introduce its fairness-aware extension, APQ-FL-Fair. Experiments show that APQ-FL and APQ-FL-Fair improve test accuracy by 4.96%–15.33% while reducing 23.14%–74.45% communication overhead compared to other methods, and exhibit stable and superior performance even under stringent latency constraints.
Xiao-Dong Li, Yulong Gao, C. Chiasserini et al.· IEEE Transactions on Cogniti...· 0 citations
Over-the-Air Computation (AirComp) Federated Learning (FL) is actively studied as a communication-efficient technique for distributed Artificial Intelligence (AI) model training. To mitigate the impact of wireless channels on the aggregated global model while addressing client energy sustainability, recent efforts have explored integrating Reconfigurable Intelligent Surfaces (RIS) and Simultaneous Wireless Information and Power Transfer (SWIPT) into AirComp FL. In this context, literature has mainly focused on radio resource allocation for optimized SWIPT and RIS-assisted Downlink (DL) model broadcasting and Uplink (UL) AirComp model aggregation. Nevertheless, existing works largely treat the communication design of AirComp FL in isolation, neglecting the tight coupling between radio and compute resource allocation. In this paper, we address this gap by modeling the radio-compute dependency in AirComp FL and optimizing harvested energy to sustain client-side local training and model transmissions. To this end, we jointly optimize the RIS configuration, SWIPT power-splitting ratio, DL transmission time, and local computing frequency to minimize the total communication and computation overhead in latency and energy. The original non-convex problem is decomposed into two independent subproblems, which are solved iteratively via a combination of low-rank optimization and min-max convex reformulation techniques. Numerical evaluations confirm that integrating RIS and SWIPT into AirComp FL leads to higher accuracy, and reduced latency and energy overheads across the FL pipeline.
Stefanos Voikos, P. Charatsaris, Maria Diamanti et al.· IEEE Transactions on Wireles...· 0 citations
Wireless multimodal federated learning (MFL) is promising for privacy-preserving edge intelligence, but its practical deployment is challenged by modality heterogeneity, client resource heterogeneity, and the high communication and computation cost of multimodal models. This paper proposes a beamforming-aided wireless MFL framework with dynamic neural network pruning, termed mFedDNP. In mFedDNP, each scheduled client trains a pruned unimodal submodel matched to its communication and computing capability, while the server uses modal compensation to handle missing modalities and maintains updates in a storage-efficient manner. We characterize the effect of pruning on communication and computation cost, and formulate a joint design over client scheduling, neural network pruning, and resource block (RB) allocation under perround latency constraints. We further develop a staleness-aware solution that combines adaptive pruning with matching-based RB allocation. Experimental results on CREMA-D and UCI-HAR show that mFedDNP consistently outperforms benchmark methods, achieving up to 4.89 percentage points higher Macro-F1 and 23.52% shorter convergence time.
Chun-Feng Xie, Zhi-Xiong Chen, Wenqiang Yi et al.· 2026 IEEE/CIC International...· 0 citations
Emerging 6G applications impose heterogeneous, time-varying requirements on both wireless spectral resources and distributed computing resources. Existing works address hybrid precoding (HP) design and computing power network (CPN) task scheduling in isolation, and predominantly rely on static, offline optimization that cannot adapt to dynamic channel fluctuations and computing load variations. More critically, no prior work enforces a unified fairness criterion that spans the communication rate region and the computing latency distribution simultaneously. In this paper, we formulate the joint HP and CPN scheduling problem as a fairness-constrained Markov decision process (MDP) and propose an Online Adaptive Actor-Critic $(\text{OA}^{2} \mathrm{C})$ algorithm to solve it in real time. ${OA}^{2} \mathrm{C}$ incorporates three novel components: (i) a dual-timescale update rule that decouples fast-varying channel adaptation from slow-varying load balancing, (ii) a heterogeneity-aware state encoder that jointly embeds channel state information, task-type profiles, and hardware capability tensors, and (iii) a Lyapunov-guided fairness regularizer that provably bounds the long-run deviation from the target $\alpha$-fair rate-latency trade-off surface. Simulation results on a spherical-wave massive MIMO channel with realistic CPN hardware heterogeneity show that OA ${ }^{2} \mathrm{C}$ achieves up to $2 2. 4 \%$ improvement in Jain's fairness index, 19.7% reduction in mean task completion latency, and throughput within 3.1% of the offline DPC-based nonlinear HP upper bound, while converging ${2. 3} \times$ faster than vanilla proximal policy optimization (PPO) baselines.
Jia-Qi Huang, Bei-Bei Zhang, Xin-Yuan Li et al.· 2026 IEEE/CIC International...· 0 citations
To mitigate the scalability bottleneck in the radio access network (RAN) in federated edge learning (FEEL), over-the-air federated learning (AirFL) exploits waveform superposition over multiple-access channels (MACs) for analog model aggregation. However, coherent AirFL typically relies on stringent PHY-layer conditions such as accurate channel state information (CSI), tight time/frequency synchronization, and frequent transceiver calibration for signal alignment. However, these requirements, if not impossible to be met, incur substantial communication and computation overhead. In this paper, we propose a non-coherent AirFL (NCAirFL) protocol over a broadband single-antenna MAC, leveraging binary dithering, unbiased non-coherent detection, and long-term error feedback to waive the need for instantaneous CSI. For NCAirFL with general smooth non-convex objectives and a constant learning rate, we establish a convergence bound achieving the convergence rate in the same order of $\mathcal{O}(1/\sqrt{T})$ as communication-ideal FedAvg, where $T$ is the total number of communication rounds. To further improve communication efficiency under data and wireless resource heterogeneity, we also derive a lower bound on the expected single-round objective decrease in the global loss conditioned on device scheduling, building upon which a surrogate objective function is obtained for jointly optimal device selection and power control. Experimental results on MNIST and CIFAR-10 corroborate that NCAirFL achieves learning performance close to FedAvg in practical settings, with the proposed device scheduling policy substantially accelerating convergence.
Hai-Feng Wen, Nicolò Michelusi, Osvaldo Simeone et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
