Federated Learning (FL) typically involves processing redundant data on resource-constrained edge devices, resulting in prolonged training time. A promising strategy to accelerate FL is on-device data sampling. SOTA methods generally select a subset based on sample importance before each local epoch. However, these methods, which employ mini-batch gradient descent on the static subset, suffer from outdated sample importance, leading to suboptimal sampling efficiency. Furthermore, these sampling methods fail to address the biased gradient expectation introduced by importance sampling, further degrading model accuracy. In this paper, we propose FedTT, a novel framework designed to accelerate FL with improved accuracy through pipelined intra-batch data sampling and model training. Specifically, at the algorithm level, to enable real-time data sampling with unbiased model updates, FedTT performs on-device sampling from the fixed-size mini-batch at each iteration and applies gradient correction to the variable-size sampled micro-batch. At the system level, to further accelerate FL training, FedTT implements a well-designed parallelism and synchronization mechanism that enables pipelined execution of intra-batch data sampling and model training on CPU-GPU architectures. Finally, we conduct extensive real-world experiments and simulations to demonstrate the effectiveness and universal adaptability of FedTT. Compared to the SOTAs, our evaluation on four datasets shows that FedTT improves time-to-accuracy performance by 1.23 × ∼ 2.51 × and model accuracy by 0.39% ∼ 11.24%, with real-world experiments further validating its real-world effectiveness by achieving a 1.75 × speedup and a 9.86% accuracy improvement. Integrated with various FL algorithms and importance criteria, FedTT consistently delivers performance gains. Code has been open sourced.

Unmanned aerial vehicles (UAVs) equipped with sensors, computing units, and communication modules, together with ground devices, constitute a ubiquitous integrated low-altitude network, which can provide users with sustainable computing and communication services in areas where terrestrial infrastructure has been compromised or rendered inoperable. Federated learning-enabled UAV (FL-UAV) wireless networks fully utilize the computational and communication capabilities of UAVs to protect user data privacy by exchanging model updates with ground devices. However, facing the challenges of low energy utilization efficiency and high training latency caused by UAV deployment, resource allocation, and communication overhead in FL-UAV. Existing solutions do not achieve efficient communication and resource scheduling to solve the energy and delay optimization issues in FL-UAV wireless networks. In this paper, we propose an air-to-ground integrated federated distillation (AirFD) framework for UAV-assisted mobile computing and communication networks, which significantly reduces communication overhead between UAV and ground devices by introducing knowledge distillation to transmit average logits instead of model parameters. Furthermore, we formulate cross-layer resource scheduling in AirFD as a non-convex optimization problem to achieve a trade-off between energy consumption and delay. To solve this nonlinear coupling and NP-complete problem, we use successive convex approximation and greedy algorithm to obtain the local optimal solution. Simulation evaluation and field experiments confirm the effectiveness of our proposed method in reducing communication costs and training delays by nearly 40%, and increasing energy utilization by about 50%.

As a computing paradigm tailored for resource-constrained client devices, federated learning based on model pruning compresses the model size by removing unimportant parameters in the neural network, which has shown outstanding results in improving model efficiency and reducing computing costs. However, previous works simply customized a unified static model pruning rate, ignoring the heterogeneous capabilities of clients and the impact of pruning on different layers of the model during continuous iteration. In this paper, we design a novel Federated learning framework based on Dynamic Layer-wise Pruning, named FedDLP, which is capable of pruning at the hierarchical level depending on the client’s capability and model similarity to improve model efficiency and maintain model performance. This framework consists of two parts. First, pre-training customizes the initial pruning rate: We set the initial pruning rate for each layer according to the different capabilities of heterogeneous clients during the pre-training stage. Second, adaptively optimize the pruning rate: We use cosine similarity to quantify the contribution of each layer of the client model to the global model, thereby adaptively and dynamically optimizing the model pruning rate. Experimental results verify that our proposed method improves model efficiency by 2 to 3.5× compared to the state-of-the-art baselines, while achieving an accuracy difference of no more than 2% compared to the unpruned model.

Asynchronous Federated Learning (AFL) is a promising technique to enable efficient and flexible distributed learning across heterogeneous devices. However, AFL faces the challenge of handling stragglers, i.e., devices that have high latency or low participation rate, which can degrade the convergence speed and accuracy of the global model. Existing solutions either discard or penalize the updates from stragglers, which may result in losing valuable information or introducing bias. In this paper, we propose FedVDA, a novel AFL framework that significantly improved the QoS of AFL in terms of model accuracy and training time cost by effectively utilizing and compensating for the updates from stragglers. Specifically, Fed-VDA introduced the dynamic window protocol that dynamically adjusts the server’s waiting time in each round based on the estimated completion time of the devices. We further designed a version control mechanism that corrects the stale gradients of the stragglers by supplementing the missing training rounds. Extensive experiments on three public datasets demonstrate that FedVDA achieves, on average, 2.7× faster convergence speed and 5.1% higher accuracy than state-of-the-art AFL methods. Moreover, we open-sourced FedVDA1and show that it is non-intrusive and highly scalable, which enables easy integration with other AFL algorithms and improves their QoS with no-pain in large-scale federated learning systems.

Deep learning-based visual localization methods use convolutional neural networks to directly regress the position of a target. However, previous studies only consider localization in a single scene and neglect personalized localization in multiple scenes. Furthermore, changes in the scene result in a reduced accuracy due to the model’s lack of adaptability. Moreover, traditional centralized training methods pose data privacy concerns. In this paper, we propose a personalized semi-supervised federated learning framework with cloud-edge-end collaboration, called FedVL. The hierarchical architecture extends single-scene localization to multiple scenes, while the federated learning mechanism ensures data privacy. In this framework, we apply personalized federated learning to achieve scene-specific model and employ semi-supervised federated learning to allow the localization model to adapt to scene changes. Experiments conducted on indoor and outdoor datasets demonstrate the effectiveness of this approach.

Personalized federated learning (PFL) allows distributed clients and the server to share model parameters instead of raw data, aiming to customize a personalized model for each client. However, the naive design of weighted aggregation in previous studies can easily transfer the deviated sample knowledge to some local models, while ignoring the implicit relationship between the model layer and the training samples that match each other. Furthermore, PFL requires frequent parameter exchange between clients and the server to accurately obtain a personalized model suitable for its needs, which further exacerbates the problem of communication overhead. In this paper, we present a PFL framework, named DhcPFL, which is featured by the fine-grained observation that the updated parameters of each layer of the model implicitly provide information about the distribution of training samples, while achieving both improvements in model performance and communication cost. We manage to do these by innovating in the following aspects. Specifically, we propose a hierarchical clustering method based on Wasserstein distance for model parameters, which can fine-grained match clients with similar individual needs through the update of parameter layers. Based on this, we further propose a novel layered aggregation rule, which allows model parameters of partial layer aggregation instead of complete model parameters to be exchanged between client and server, effectively alleviating the problem of excessive communication overhead. Experiments on three public datasets demonstrate that our proposed method achieves about $2 \%$ performance improvement and $3.5 \times$ communication efficiency compared to the current baseline.

Air access networks have been recognized as a significant driver of various Internet of Things (IoT) services and applications. In particular, the aerial computing network infrastructure centered on the Internet of Drones has set off a new revolution in automatic image recognition. This emerging technology relies on sharing ground-truth-labeled data between unmanned aerial vehicle (UAV) swarms to train a high-quality automatic image recognition model. However, such an approach will bring data privacy and data availability challenges. To address these issues, we first present a semisupervised federated learning (SSFL) framework for privacy-preserving UAV image recognition. Specifically, we propose a model parameter mixing strategy to improve the naive combination of federated learning and semisupervised learning methods under two realistic scenarios (labels-at-client and labels-at-server), which is referred to as federated mixing (FedMix). Furthermore, there are significant differences in the number, features, and distribution of local data collected by UAVs using different camera modules in different environments, i.e., statistical heterogeneity. To alleviate the statistical heterogeneity problem, we propose an aggregation rule based on the frequency of the client’s participation in training, namely, the FedFreq aggregation rule, which can adjust the weight of the corresponding local model according to its frequency. Numerical results demonstrate that the performance of our proposed method is significantly better than those of the current baseline and is robust to different non-independent and identically distributed(IID) levels of client data.