To address the critical security challenges posed by the inherent broadcasting nature and heterogeneous service demands in cognitive satellite-terrestrial networks (CSTN), this paper presents a groundbreaking framework that integrates reconfigurable intelligent surfaces (RIS) with sparse code multiple access (SCMA). This framework aims to maximize the achievable secrecy rate by jointly optimizing transmit beamforming, the RIS reflection matrix, and SCMA codebook configurations, while adhering to power constraints and the quality-of-service requirements of legitimate users. To tackle the non-convex optimization problem in complex environments, we develop an intelligent decision-making mechanism based on a modified multi-agent two-delay deep deterministic (MMTD3) algorithm, which introduces a breakthrough mechanism by decoupling continuous beam control from discrete codebook selection, offering a new paradigm for AI-driven cross-domain security optimization in CSTN. Simulation results demonstrate that the proposed framework significantly outperforms existing benchmarks, verifying its potential in supporting wireless endogenous security and meeting massive heterogeneous service demands in CSTN.

Cell-free massive MIMO (CF mMIMO) systems overcome inter-cell interference, enhancing overall communication rates for next-generation networks. However, the pilot contamination exacerbates channel estimation errors and the complex connectivity makes it difficult to deal with resource allocation optimization problem. In this paper, we investigate the robust beamforming problem under channel uncertainty with the goal of improving the minimum quantile rate. Specifically, we introduce hypergraph neural network (HGNN) into the wireless resource allocation of CF mMIMO ststems for the first time, leveraging hypergraph modeling to capture the many-to-many relationships between Access Points (APs) and User Equipments (UEs). Furthermore, we significantly reduce the search space of the optimization problem by applying optimal interference suppression beamforming theory. In order to soften the sorting process, we adopt the Monte Carlo sampling strategy for data augmentation. Simulation results demonstrate that the proposed algorithm outperforms conventional schemes, achieving 14.1% performance gain and converging more than twice as fast as the state-of-the-art machine learning models.

Constructing digital twin (DT) models of user equipments (UEs) efficiently is essential for enabling real-time monitoring of UEs, providing crucial support for optimizing the operation of Social Internet of Things (SIoT) systems. However, UE heterogeneity and UE mobility concerns impede the DT deployment in SIoT. In this article, we propose a real-time DT deployment (RDTD) scheme for SIoT systems, where the heterogeneous DT modeling and the DT migration are achieved based on personalized federated learning (PFL) ideas. Specifically, we decompose the DT model into the global generalization layers and the personalization layers, based on which we propose a hierarchical PFL (HPFL)-based DT model construction mechanism. The mechanism constructs customized DT models for heterogeneous UEs through a two-stage model parameter update process, involving end-edge-center collaboration training of all parameters and fine-tuning of the personalization layer parameters. Second, based on the above mechanism for DT model construction, a low-latency DT model parameter migration algorithm is proposed. This algorithm ensures real-time interaction by migrating only the personalized layer parameters of the DT model and reconstructing the DT model. Lastly, numerical experiments verify the effectiveness of the RDTD scheme, improving modeling accuracy by 13.91%, 41.06%, and 135.35% compared to the three baselines. Additionally, our proposed scheme significantly reduces interaction latency by 29.93% compared to the baseline.