Federated Learning (FL) enables decentralized model training while preserving data privacy, however, real-world deployments are often constrained by Periodic Deterministic Communication (PDC) schedules, where communication between clients and the central server occurs at fixed intervals due to bandwidth limitations, energy constraints, or regulatory restrictions. These rigid schedules introduce fundamental challenges, including delayed model updates, model drift, inefficient convergence, and heightened sensitivity to non-IID data distributions, which undermine FL performance in practical settings. To address these limitations, we propose Federated Nested Quasi-Newton Optimization (FedNQN), a novel framework that accelerates convergence and enhances FL robustness under PDC constraints. FedNQN integrates curvature-aware central acceleration with variance-controlled local adaptation, ensuring stable learning dynamics despite restricted communication. At the global level, second-order curvature information accelerates model updates, compensating for infrequent synchronization, while local updates leverage variance-controlled optimizations to mitigate drift and adapt to heterogeneous data distributions. This coordinated optimization strategy enhances convergence speed, improves model accuracy, and maintains computational efficiency, making FL more adaptable to real-world constraints. Extensive experiments on benchmark datasets validate FedNQN’s effectiveness, demonstrating superior performance over state-of-the-art FL methods in terms of stability, scalability, and resilience to communication inefficiencies.

Federated learning (FL) in Intelligent Internet of Things (IIoT) environments faces critical challenges, including sparse client participation, non-IID local data distributions, and unreliable communication, which lead to slow convergence and high variance in global updates. To address these issues, we propose adaptive central federated momentum optimization (ACFMO), an optimization framework that enhances FL efficiency and stability under constrained participation. ACFMO integrates an adaptive central acceleration mechanism that dynamically adjusts momentum updates based on real-time client availability, preventing instability and ensuring smoother global model updates. Additionally, a variance-controlled local updating strategy refines client contributions, mitigating high variance caused by infrequent and heterogeneous updates. Extensive experiments across diverse FL scenarios demonstrate that ACFMO significantly accelerates convergence, reduces communication overhead, and improves model stability compared to state-of-the-art FL methods, making it particularly well-suited for real-world IIoT deployments where network and computational resources are constrained.

In the rapidly advancing ubiquitous intelligence society, the role of data as a valuable resource has become paramount. As a result, there is a growing need for the development of autonomous economic agents (AEAs) capable of intelligently and autonomously trading data. These AEAs are responsible for acquiring, processing, and selling data to entities such as software companies. To ensure optimal profitability, an intelligent AEA must carefully allocate its portfolio, relying on accurate return estimation and well-designed models. However, a significant challenge arises due to the sensitive and confidential nature of data trading. Each AEA possesses only limited local information, which may not be sufficient for training a robust and effective portfolio allocation model. To address this limitation, we propose a novel data trading market where AEAs exclusively possess local market information. To overcome the information constraint, AEAs employ federated learning (FL) that allows multiple AEAs to jointly train a model capable of generating promising portfolio allocations for multiple data products. To account for the dynamic and ever-changing revenue returns, we introduce an integration of the histogram of oriented gradients (HoGs) with the discrete wavelet transformation (DWT). This innovative combination serves to redefine the representation of local market information to effectively handle the inherent nonstationarity of revenue patterns associated with data products. Furthermore, we leverage the transform domain of local model drifts in the global model update process, effectively reducing the communication burden and significantly improving training efficiency. Through simulations, we provide compelling evidence that our proposed schemes deliver superior performance across multiple evaluation metrics, including test loss, cumulative return, portfolio risk, and Sharpe ratio.

We consider a setting engaging in collaborative learning with other machines where each individual machine has its own interests. How to effectively collaborate among machines with diverse requirements to maximize the profits of each participant poses a challenge in federated learning (FL). Our studies are motivated by the observation that in FL the global model attempts to acquire knowledge from each individual machine, while aggregating all local models into one optimal solution may not be desirable for some machines. To effectively leverage the knowledge of others while obtaining the customized solution for individual machine, we propose the accelerated federated training procedures with diversified global models. Based on the federated stochastic variance reduced gradient (FSVRG) framework, we propose the model-based grouping mechanism with adaptive central acceleration (MA-FSVRG) and gradients-based grouping mechanism with adaptive central acceleration (GA-FSVRG) to tackle the challenges of heterogeneous demands. The simulation results demonstrate the advantages of the proposed MA-FSVRG and GA-FSVRG over the state-of-the-art FL baselines. MA-FSVRG exhibits greater stability in performance and significant cost savings in local computation expenses compared to GA-FSVRG. On the other hand, GA-FSVRG attains higher test accuracy and faster convergence speed, particularly in scenarios with limited individual machine participation.

In the realm of healthcare data analysis, privacy concerns have been tackled by the federated learning (FL) framework. However, in the situation that heterogeneous healthcare Internet of Things (IoT) devices collect different types of data, applying FL becomes difficult. To train a model leveraging diverse healthcare IoT devices, we propose an advanced collaborative learning framework to fill the gap. With the proposed collaborative learning framework, individual IoT devices project their sensed features into a carefully developed latent space, which are transmitted to a central server. For privacy preservation, the latent local features are encoded within this space, while the samples’ labels remain securely stored in the individual IoT devices. Collaboratively, the deep neural network model is trained by both the central server and the diverse IoT devices. The central server handles the computationally intensive training processes, while the individual IoT devices evaluate the model’s performance and initiate back-propagation based on their locally stored labels. Experimental results demonstrate that the proposed collaborative learning framework achieves performance similar to centralized training and significantly outperforms individual training while preserving data privacy.

Federated learning (FL) deployed in the edge network environment is a promising approach for combining the separated training results based on the isolated local data sensed by various Internet of Things (IoT) devices. However, the limited computing resources for the training of various application models in each edge server and the communication burden among the edge server and numerous IoT devices greatly impact the realization of IoT intelligence. In this article, we propose transform-domain FL schemes based on discrete cosine transform (DCT-FA) and discrete wavelet transform (DWT-FA) to achieve better training efficiency and reduce the communication burden for IoT devices. Furthermore, when the amount of training data is limited, we propose to combine time-domain features and frequency-domain features in FL (CDCT-FA) that turns out to achieve much higher test accuracy. From the experimental results, the transform-domain FL schemes are shown to be promising, given the different constraints and requirements of various IoT intelligence applications.

Edge caching is promising to alleviate the backhaul pressure and provide low latency delivery for delay sensitive applications. However, it encounters great challenges to make adaptive content placement decisions according to the scattered explicit feedback with spatial and temporal dynamics. We propose a hybrid learning framework to obtain a more accurate prediction of users' preference by combining historical data from the central cloud and real-time data in edge networks. Two hybrid-learning algorithms, i.e., Hybrid Learning based on Alternating Least Squares (HLALS) and Hybrid Learning based on Conjugate Gradient Descent (HLCGD) are designed to achieve efficient caching decisions, where HLCGD is more efficient than HLALS at the expense of complexity. Simulation results show that, compared to the popular stochastic gradient descent strategy, the proposed algorithms can achieve superior performance thanks to more accurate prediction of users preference.

Given the highly dynamic traffic loads of mobile Internet of Things (IoT) devices and their stringent quality-ofservice requirements, i.e., access delay particularly, as well as the heterogeneous infrastructures among IoT networks, it is a nontrivial task to efficiently deploy cloudlets among large number of access points (APs) in IoT networks, especially for the access delay and network reliability, since different placement schemes would produce various network performances. To combat this issue, we are motivated to investigate in details the optimal placement of cloudlets to minimize the average access delay by applying software-defined networking (SDN) techniques to provide flexible and programmable management for cloudlets deployment in IoT networks with considering the complicated queuing process at numerous SDN-based APs. An enumerationbased optimal placement algorithm (EOPA) is first proposed as benchmark. Then we propose a ranking-based near-optimal placement algorithm (RNOPA) which is able to dynamically adapt to mobile IoT devices and their traffic loads, by treating each AP as a single server queue and adopting an efficient ranking mechanism. As corroborated by extensive simulation results, RNOPA reports access delay very close to that of EOPA. Note that RNOPA outperforms the famous K-medians clustering algorithm (KMCA) in both of average cloudlet access delay and reliability, while at the cost of a much lower computational complexity than KMCA.

Leveraging the concept of software-defined network (SDN), the integration of terrestrial 5G and satellite networks brings us lots of benefits. The placement problem of controllers and satellite gateways is of fundamental importance for design of such SDN-enabled integrated network, especially, for the network reliability and latency, since different placement schemes would produce various network performances. To the best of our knowledge, it is an entirely new problem. Toward this end, in this paper, we first explore the satellite gateway placement problem to obtain the minimum average latency. A simulated annealing based approximate solution (SAA), is developed for this problem, which is able to achieve a near-optimal latency. Based on the analysis of latency, we further investigate a more challenging problem, i.e., the joint placement of controllers and gateways, for the maximum network reliability while satisfying the latency constraint. A simulated annealing and clustering hybrid algorithm (SACA) is proposed to solve this problem. Extensive experiments based on real world online network topologies have been conducted and as validated by our numerical results, enumeration algorithms are able to produce optimal results but having extremely long running time, while SAA and SACA can achieve approximate optimal performances with much lower computational complexity.

Mobile edge computing (MEC), as an extension of the cloud computing paradigm to the edge network, is a promising solution to provide resource-intensive and time-critical applications to mobile users. It overcomes some obstacles of traditional mobile cloud computing by offering ultra-short latency and less core network traffic. This paper proposes a new framework based on the architecture of MEC to deliver cloud services to the edge. We introduce enumeration based optimal placement algorithm (EOPA) and divide-and- conquer based near-optimal placement algorithm (DCNOPA) to attain minimal data traffic by distributing virtual machine replica copies (VRCs) of applications to the edge network. Simulation results show that compared to the famous K-medians clustering algorithm (KMCA), the performance of DCNOPA is much closer to that of EOPA with lower computational complexity. Furthermore, we investigate the optimal number of VRCs within a given limitation of benefit-to-cost ratio.