Federated learning (FL) enables collaborative model training across distributed data sources through local computations and aggregation of model parameters, thereby preserving data privacy. Federated split learning (FSL) extends this paradigm by partitioning deep neural networks and transmitting only intermediate feature representations instead of raw data, thus offering dual-layered privacy protection while mitigating computational bottlenecks. However, identifying optimal split points in dynamic and heterogeneous environments to balance model performance, privacy preservation, and computational efficiency remains a significant challenge. To address these limitations, this paper proposes an adaptive state augmentation method based on retrieval-augmented generation (RAG). Our method employs dynamic k-values and relevance scoring to retrieve key historical states from a memory repository, fuses current and historical features using an hourglass encoder and soft attention mechanisms, and generates high-fidelity augmented states via a residual-derived decision module. The resulting augmented states incorporate historical contextual insights, enhancing the agent’s environmental awareness and enabling more precise and reliable split decisions. Empirical evaluations demonstrate that, compared to conventional reinforcement learning techniques, the proposed method maintains model accuracy while reducing average training time by 25%–40% and significantly mitigating load imbalances and vulnerability to data leakage.

For future Internet of Vehicles (IoV), communications and computing will converge to provide services. Federated learning (FL), as one of the typical distributed computing technologies, needs to be integrated with IoV. For such integration, FL suffers from the straggler effect that the entire learning speed is lowered down, because of the existence of the devices, such as low-powered road side units and vehicles, taking more time to complete their tasks. Although the existing mechanisms reduce straggler effects by adopting asynchronous mechanisms and clustering mechanisms, they lack the detailed analysis of the reasons and the impacts of each cause, leading to inefficiencies in the design of algorithm. Additionally, most of the existing work only considered the impact of a single factor in computation, communication, or data distribution, which lacks comprehensive on research for causes of stragglers effects. The bottleneck is that it is laborious to observe the time delay precisely with the existing high-calculating evaluations. In this article, we elaborately explore the effects of computing power, communication capability, and data distributions on the straggler effects with carefully designing and conducting the extensive experiments. After investigations, we propose a novel learning completion time estimation formula for low computing capability devices with mini-batch stochastic gradient decent (SGD). We compare our proposed estimation formula with the one based on floating operation per second (FLOPs). Through the evaluations, our formula can demonstrate the improvement up to 72.4% at docker and 32.4% at Raspberry Pi device compared to the existing work.

Model partitioning techniques, which decompose and collaboratively execute subtasks of deep neural networks (DNNs), have emerged as a critical strategy for enhancing distributed inference efficiency. However, in mobile edge computing (MEC), dynamic load fluctuations at edge nodes and the complexity of cross-node task dependencies make the delay minimization problem extremely challenging. Existing studies predominantly adopt a decoupled optimization framework that separately addresses partitioning-offloading and pipeline scheduling, neglecting their inherent cyclic state-dependent coupling. This oversight leads to suboptimal solutions, such as pipeline stagnation caused by mismatched computation and communication timestamps. To address these challenges, we propose a multi-endpoint directed acyclic graph (DAG)-driven cooperative optimization approach, enabling partitioning-offloading and pipeline scheduling in MEC. Specifically, the approach involves two core steps: 1)Dynamic pre-scheduling: We propose an improved DNN scheduling algorithm for constrained subtasks, which simulates node-level queuing delays and pipeline stalls under real-world constraints, translating runtime states into latency objectives. 2)Partitioning and offloading solution retrieval: Based on latency objectives, we introduce a novel multi-endpoint DAG structure and design a multi-node collaborative optimization retrieval algorithm, enabling adaptive partitioning-offloading remapping of subtasks. Experiments demonstrate the superiority of the proposed method over other advanced methods, reducing the time overhead by an average of 24% and 75% in two different scenarios, respectively. The resource code can be found at: https://github.com/aiheiheiheii/Partition_Scheduling.git.

Recent research on unmanned aerial vehicle-based (UAV) computational offloading algorithms has employed energy harvesting mechanisms to improve the efficiency of edge computing. However, these approaches treat UAVs as relays or power providers which are rarely regarded as the computational nodes in the energy harvesting condition. The most challenging issue is not only the computational efficiency, but also the energy stability of individual computing devices. Stable energy state represents a stable computing service capability, which is very important for edge systems. It depends heavily on proper computational offloading algorithms. In this study, we construct a novel model that uses a cluster of UAVs with energy harvesting capability as a computational core. It is capable of providing long-term computational services for various scenarios. Then, we construct a Lyapunov function through a designed virtual battery energy queue and prove the existence of an upper bound for the Lyapunov drift-plus-penalty function through mathematical transformations. Therefore, we obtain a theoretically stable battery energy queue and design a Lyapunov-chain offloading algorithm based on it. Simulation results show that the proposed Lyapunov-chain offloading algorithm is able to maintain the strong energy stability of each node. It also provides robustness for edge UAV clusters while minimizing the execution delay compared to the baseline offloading scheme.

Owing to the technology of 5G and beyond, collaborative edge computing-as-a-service has enabled trillions of interconnected edge applications. It has also become a prospective paradigm for providing computing services by offloading computationally intensive assignments to mobile-edge servers or fog nodes due to terminals constrained computing and caching resources. Nevertheless, in this process, trust of computing-as-a-service scheduling and edge data sharing in heterogeneous systems is an unavoidable challenge of paramount importance. As a powerful tool that addresses security issues, blockchains can ensure the trustworthiness and irreversibility of computing data by consensus mechanisms. However, in the Industrial Internet of Energy (IIoE), the storage burden of a single blockchain has increased. Therefore, from the perspective of a stable real-time operation, we propose a multiedgechain structure that accommodates thousands of edge data and promotes on-chain data efficiency to achieve cross-chain edge data sharing for heterogeneous blockchain systems. Moreover, aiming at the profits of computing resource scheduling in the IIoE, a two-stage Stackelberg game strategy with an optimal scheduling demand and reward is provided considering the edge user’s preferences and risk factors. Finally, the simulation results verify the superiority of the proposed scheme, regarding the game equilibrium, utility optimization, and data sharing efficiency of cloud–edge collaboration.