2026· IEEE Transactions on Wireless Communications· Vol 25, pp. 19226-19242· 0 citations· 58 references
Computer Science
Abstract
Integrating over-the-air computations into the model aggregation process of federated learning (FL) offers a promising solution to mitigate the communication bottleneck in FL model training. In this approach, all the clients modulate their intermediate parameters, such as gradients, onto the same set of orthogonal waveforms and transmit the resulting signals to the edge server simultaneously. Capitalizing on the superposition property of the radio channel, the server can extract an automatically aggregated global gradient from the received radio signal. However, the limited number of orthogonal waveforms imposes a constraint on the dimensionality of transmittable updates, hindering the adoption of more advanced, but high-dimensional models. In light of this challenge, we propose OFLight, a lightweight, yet effective, gradient compression algorithm tailored for OTA-FL systems. Specifically, in each communication round, the edge server constructs a low-rank projection matrix based on the received gradient matrix from the previous round (initialized with an independent and identically distributed standard normal matrix in the first round) and broadcasts it, along with the global model, to all clients in the system. Based on this matrix, every client projects its locally updated gradient matrix into a low-dimensional subspace through a linear operation. The clients upload only their compressed gradients via OTA computations, and the edge server can perform a linear decompression on the received signal, retrieving the original gradient dimension. Moreover, an error feedback mechanism is incorporated to compensate for the approximation error under aggressive compression. We derive analytical expressions for the convergence rate of both convex and non-convex loss functions, quantitatively demonstrating the effect of OFLight on the OTA-FL training efficiency. We also conduct extensive experiments to corroborate the efficacy of the proposed method.
Federated learning (FL) applications normally employ large deep learning (DL) models, resulting in excessive communication overhead in the deployment of FL over resource-constraint mobile edge networks. To achieve better scalability for DL-based FL, we capitalize on both the asymmetric nature of mobile networks and the distinct effects of partial transmissions on FL training for the global and local models. We propose Fed-DynAmal, an FL framework that decreases the number of parameters transmitted in the uplink (clients-to-server) while concurrently achieving better model performance. The underlying idea is that each selected client sends a partial DL model to the server by omitting several sub-blocks from the trained local model. Crucially, we drop the assumption that transmitted local models can still be used for inference, thereby allowing for greater model variability. At the server, we introduce amalgamation, a process to merge different partial local models into an inference-viable full model. Essentially, amalgamation is a bridge for performing aggregation at the sub-block level. Interestingly, as the key takeaway, communication efficiency versus model performance is not necessarily a trade-off in FL: Our extensive experiments show that Fed-DynAmal can effectively improve communication efficiency while still concurrently achieving higher accuracy and enhanced robustness.
Zihan Chen, H. Yang, Tony Q. S. Quek et al.· IEEE Transactions on Cogniti...· 0 citations
Nowadays, split federated learning (SFL) has emerged as an effective paradigm for enabling privacy-preserving collaborative intelligence across heterogeneous devices with limited computation. However, SFL incurs significant communication overhead in wireless networks due to the uplink transmission of high-dimensional smashed data, which degrades network efficiency. To mitigate the communication bottleneck, we propose a prototype-based SFL framework ProtoSFL. Specifically, each selected client computes local prototypes for observed classes and uploads them to the server. Based on the received prototypes, the server derives global prototypes and optimizes a weighted objective that combines classification loss with prototype alignment loss. The server then updates the model accordingly and returns personalized prototype gradients to the clients. Simulation results verify the effectiveness of ProtoSFL in reducing communication overhead, achieving a substantial reduction in uplink communication, while maintaining competitive testing accuracy under various heterogeneous data settings compared with SFL baselines.
The rapid growth of sensor-equipped edge devices calls for machine learning paradigms that are both lightweight and privacy-preserving. Federated learning (FL) enables distributed AI on edge data but still demands heavy forward and backward propagation on resource-constrained sensors. Split learning (SL) offloads computation to a server, yet its intermediate activations can leak private information. Homomorphic encryption closes this gap at the cost of prohibitive computation and bloated communication. We propose HFL-Lite, a hierarchical federated learning framework that achieves practical privacy without cryptographic primitives. Each sensor encodes its raw data into a compact sparse feature code via a lightweight sensor encoder. Only this code, which is insufficient to reconstruct the original data, leaves the device. An edge server synthesizes surrogate training samples from the uploaded codes, trains a teacher model, and distills knowledge back to the sensors. Because the sensors perform only one-way uploads, backward propagation, gradient transmission, and plaintext activation leakage are all eliminated. Experiments on the UCI HAR dataset show that HFL-Lite reduces per-round communication to 0.5 KB and sensor-side latency to 9 ms while delivering competitive accuracy.
Cangming Liang, Kuan Ching Li, Zulong Diao et al.· Tsinghua Science and Technol...· 0 citations
Over-the-air computation-assisted federated learning (OTA-FL) exploits the superposition property of the wireless channel to markedly reduce the latency and bandwidth requirements of federated learning. Devices adapt their transmit power to enable over-the-air aggregation of the local models. However, imperfect channel state information (CSI) and power constraints distort the aggregated model update at the receiver and affect the learning algorithm. We provide a novel, comprehensive analysis of OTA-FL schemes, which encompasses scaled-down channel inversion (SCI), truncated channel inversion (TCI), and controlled descent algorithm (CDA). Unlike prior studies that assume bounded estimation errors, we study a more realistic model in which the estimation error has unbounded support. Our analysis addresses both fixed and adaptive learning rates. While prior works on imperfect CSI focus only on convergence in expectation for a specific scheme and assume fixed learning rates, we provide technically stronger almost sure convergence guarantees for multiple schemes when the number of devices is large. Extensive experiments on linear regression and CIFAR-10 classification validate our theory even for a small number of devices. Adapting the learning rate leads to convergence even with very noisy estimates.
S. Adhikary, N. Mehta· IEEE Transactions on Wireles...· 0 citations
Federated fine-tuning is bottlenecked by communication: FedAvg and pseudo-gradient schemes transmit a payload that scales with the model, and gradient compression shrinks it by only a constant factor. We take a different lever. Mapping networks generate a network's weights from a small trainable latent through a frozen affine projection; because the map is shared and affine, averaging latents is exactly averaging the generated weights. We turn this into a practical low-bandwidth federated channel with two changes: a low-rank, seed-regenerable factorisation of the projection (cutting generator memory from ~80 GB to ~10 MB), and a delta formulation $\theta = \theta^{\mathrm{pre}} + U V^{\top} z$ that learns an additive correction around a shared centrally-pretrained base -- federated fine-tuning, which is what makes the method work at scale. A frozen orthogonal classifier head further removes the head from the payload while improving accuracy. On CIFAR-100 with ResNet-18+GroupNorm, our method (FLITE, Federated Low-rank Iterative Training Engine) communicates 1,280 floats (~5 KB) per client per round -- an 8718x reduction -- and reaches 74.67%, within ~0.5 pp of full-weight FedAvg. The averaging identity holds to floating-point precision ($6 \times 10^{-8}$); the method sits one to two orders of magnitude below PowerSGD and top-k on the bandwidth-accuracy Pareto; it matches or exceeds full-weight FedAvg under strong non-IID skew. int4 latents reach 648 bytes per round at unchanged accuracy, whereas int4 full-weight FedAvg collapses to chance.
Experimental results on heterogeneous MNIST and CIFAR-10 settings show that QEF-GT-AdamW consistently improves robustness and convergence performance over representative DecL baselines while achieving favorable accuracy-communication trade-offs under limited wireless resources.
Thieu Van Nguyen, T. Nguyen, Ons Aouedi et al.· 0 citations