With the high penetration rate of distributed renewable energy, the data gathering frequency and access latency demand in the new power distribution services have increased rapidly. This has placed higher requirements on the resource-constrained Wireless Sensor Networks (WSN) routing in terms of latency. Traditional centralized resource allocation mechanisms depend on global real-time state collection and struggle to satisfy stochastic low-latency service demands. Therefore, we propose a distributed asynchronous routing algorithm for power distribution WSN adapted to multiple latency-sensitive service data access. A collaborative allocation system model for heterogeneous communication resources of WSN is constructed, and the original optimization problem is decoupled into multiple layers. The distributed cooperative scheduling mechanism for bandwidth allocation and deep reinforcement learning is employed to solve the problems in these layers. Simulation results show that the proposed algorithm enables distributed asynchronous cooperation among nodes based on incomplete network state information, improves the minimum ratio of latency redundancy of services by approximately 20%, significantly enhancing latency-guaranteed access capability.
Radio maps provide the spatial distribution of large-scale radio signal power in specific geographic areas, and are widely used in fields such as resource allocation optimization and network planning design. However, current research treats base station attributes as implicit static contextual information and learns the statistical correlation between environmental geometry and signal distribution. This simplification limits these methods to single-base station (SBS) scenarios, making them difficult to capture the underlying physical characteristics of radio propagation in real-world environments, which are essentially complex communication systems with multi-base station (MBS) and multi-antenna type (MAT) heterogeneity. To address this challenge, we utilize an attention mechanism to construct a dynamic interference prediction mechanism and propose a novel physics-conditioned radio map estimation (RME) model, named the Multimodal Knowledge-Guided Multi-Base station U-shaped Radio Network (MMKG-MBURNet). Firstly, an efficient multimodal fusion framework is proposed to integrate the physical parameters of different base stations, enabling it to learn the complex dynamics of multi-source signal superposition and capture the underlying laws of propagation in MBS and MAT scenarios. Then, the Agent Latent Attention (ALA) mechanism based on low-rank approximation is proposed to address the computational bottleneck in large-scale data scenarios. Finally, a simulation dataset that is closer to real-world parameter setups is constructed, providing a solid data foundation for model validation. Experiments show that MMKG-MBURNet exhibits superior accuracy and robustness across multiple datasets, demonstrating its practicality and effectiveness.
Low Earth Orbit (LEO) satellite networks provide global low-latency connectivity, but their predictable topology and broad access surface make them vulnerable to Link Flooding Attacks (LFAs). Due to constrained satellite resources and observable orbital states, existing defenses such as topology obfuscation and traffic scrubbing are difficult to deploy directly. This paper proposes a game-theoretic hierarchical bandwidth allocation framework for LFA-resilient LEO networks. The data delivery process is divided into a Demand Access Subgame and a Demand Transmission Subgame, which model uplink access control and inter-satellite link/downlink resource allocation, respectively. Satellites and links announce prices according to resource states, while users determine bandwidth requests based on service requirements and prices. Differentiated pricing is further combined with a history-aware reputation coefficient to adjust resource allocation priorities under congestion. We derive users’ best responses, prove the existence of equilibria in the hierarchical game, and develop iterative algorithms for pricing and bandwidth allocation. Finally, simulations are conducted to evaluate the proposed bandwidth allocation scheme in terms of performance and effectiveness against LFAs.
Data-driven decision-making is a key enabler of intelligent enterprises, yet centralized machine learning (ML) poses privacy and efficiency risks. Federated learning (FL) offers a decentralized solution. Still, it faces (1) high communication cost, (2) vulnerability to poisoning and inference attacks, and (3) degraded performance on non-independent and identically distributed (non-IID) data. This paper presents FEDANIL+++, a privacy-preserving and communication-efficient FL model with three phases: (i) dynamic gradient compression, (ii) multi-layer protection via cosine similarity (CS) filtering, CKKS-FHE, and blockchain aggregation, and (iii) personalized FL (PFL) through clustering and adversarial learning. Experiments on FEMNIST, CIFAR-10, and TON-IoT show that FEDANIL+++ boosts accuracy by 15.75%, cuts communication by 25.87%, and reduces computation by 4.5% over baselines. Privacy metrics, Kullback–Leibler (KL) divergence, Shannon entropy, and correlation analysis confirm that CKKS-FHE effectively conceals sensitive patterns. The implementation of FedAnil+++ is publicly available on GitHub1.
Accurate and fast beam alignment is crucial in millimeter-wave (mmWave) systems, which utilize large-scale antenna arrays to resist harsh propagation environments. Although existing deep learning-based methods provide feasible solutions for beam alignment, their adaptability to environmental variations is constrained because their effectiveness relies on the consistency of the data distributions between the training and testing sets. In this paper, we propose MTSBP, a novel meta-learning-based method to obtain a generalizable beam predictor. Specifically, we introduce a trajectory-shifting mechanism into vanilla Model-Agnostic Meta-Learning (MAML). This mechanism shifts both the inner parameters and the meta parameters during the learning process to correct task-specific learning trajectories in a more consistent manner. This design helps the model learn the underlying task distribution and enhances its adaptability, thereby enabling large-step gradient adaptation to new tasks using a few new samples. Our experimental results show that MTSBP demonstrates a strong capability in addressing the beam prediction problem across various scenarios. Specifically, in the NLOS-to-LOS adaptation case, MTSBP delivers a 32.81% accuracy gain and a 4.31 bits/s/Hz spectral efficiency gain with 16 probing beams, surpassing the state-of-the-art method.
