Speech Emotion Recognition (SER) in real-world multimodal settings is challenged by non-stationary acoustic noise, modality redundancy, and annotation uncertainty, which often lead to unstable representations and performance degradation. Existing approaches lack structural modeling mechanisms to ensure robustness under noise and modality perturbation. We propose GS3-ICL, a graph-guided framework that integrates structured representation induction, sparse token selection, and invariance-consistent learning. The model constructs modality-specific graphs to enforce structural smoothness, performs graph-constrained discrete subset selection to retain topology-coherent tokens, and adopts perturbation-consistent training to enhance stability under noisy and missing-modality conditions. Extensive experiments on four benchmarks demonstrate improvements over prior methods. On IEMOCAP, GS3-ICL achieves 78.92 WA and 79.88 UA. It further obtains 67.55 WF1 on MELD, and 84.64 WA on SAVEE. The model further demonstrates controlled and gradual performance degradation under severe acoustic perturbations, indicating robustness in noisy real-world scenarios. These experimental results fully validate the efficacy of the proposed graph-guided structural learning paradigm for constructing reliable, robust, and generalizable multimodal emotion recognition systems.

Satellite communication links leveraging multi-source physical layer features provide critical authentication technology for secure satellite networks. However, existing physical layer security schemes suffer two major limitations: i) most methods perform one-shot decisions without tracking trust evolution over changing link geometries; ii) fixed decision rules fail to accumulate evidence and degrade robustness when spoofing parameters adapt. We propose a multi-stage dynamic trust authentication framework. Firstly, we introduce a multi-feature statistical detection module that performs real-time preliminary authentication based on Doppler shifts and power residual analysis. Then we design a particle-filtering-based trust-state estimation mechanism that fuses historical credibility with new observations via adaptive resampling, enabling continuous trust evolution and swift anomaly response. Experiments on STK-simulated datasets demonstrate that the proposed method increases secure throughput by 5.5% and reduces false-alarm and missed-detection rates by 27% and 29%, respectively.

Cooperative path planning for UAVs using hierarchical optimal control is a critical technology for urban emergency communication networks. Existing methods typically decouple trajectory optimization from communication scheduling, which inherently leads to suboptimal performance. Attempting to solve the problem monolithically creates a large-scale Mixed-Integer Non-Linear Program that is computationally intractable for real-time deployment. Compounding this, current models often overlook the need for dynamic UAV role-switching, limiting the system’s functional flexibility and operational adaptability. We propose a hierarchical control framework that decomposes the problem into a dual-layer model predictive control architecture. Specifically, we linearize the original problem within a receding horizon control loop for long-term task and trajectory planning in the strategic layer. Guided by this, the lower trajectory layer employs non-linear model predictive control, solved via sequential convex programming, to generate real-time trajectories and perform adaptive role switching. Simulation results demonstrate that the proposed framework reduces mission makespan by 8.6% and increases the total data collected by 2.4%, validating its efficiency and robustness.