Disaggregated 5G/O-RAN edge multicast spans multiple vendors and administrative domains, creating a need for auditable trust signals that do not expose sensitive evidence or disrupt real-time media delivery. This paper presents ZK-TrustLLM, a three-plane architecture that preserves the real-time DTLS/RTP multicast media path while separating evidence retention from trust verification. The evidence plane stores audit artefacts off-chain using the InterPlanetary File System (IPFS) via content addressing (CIDs), while the trust plane records commitments, proof outcomes, and policy-triggering events on a permissioned blockchain. Zero-knowledge submissions prove that a reported score lies within an admissible range and is cryptographically bound to an evidence commitment, while raw evidence remains off-chain for authorised audit. Trust outputs can then drive multicast control policies such as Logical Key Hierarchy (LKH) rekeying or subgroup isolation. We implement and evaluate a 10-tenant prototype (50 submissions) on a local chain with co-located IPFS, observing predictable confirmation latency (p50=6 ms, p95=7 ms) with stable gas (mean ≈ 137,783). The released artefact includes executable baselines and NDJSON logs to support reproducible evaluation.
Fluid antenna (FA) has emerged as a highly promising candidate in wireless communications, attributed to its exceptional design flexibility and enhanced degrees of freedom (DoFs). This paper addresses the challenge of direction-of-arrival (DoA) estimation for closely spaced signals under low signal-to-noise ratio (SNR) and limited snapshot conditions. To tackle this issue, a dual-driven FA system is proposed, which synergistically integrates data-driven and model-driven paradigms. Drawing inspiration from the Toeplitz algorithm, a deep neural network (DNN)-based multi-layer perceptron (MLP) architecture is developed to learn the mapping relationship between input covariance matrices and labeled covariance matrices. To ensure the output covariance matrices are compatible with subsequent subspace-based search procedures, a constrained forward computation function is designed. Furthermore, a dual-constrained loss function is formulated to balance the numerical precision and structural consistency of the output covariance matrices. Simulation results demonstrate that the proposed method achieves significant performance improvements over conventional DoA estimation algorithms in terms of estimation accuracy.
To democratize zero-touch network automation, we need to connect natural language user intents with formal, standards-compliant specifications. Large Language Models (LLMs) are promising but unreliable. We propose a hybrid AI-symbolic framework that uses LLMs for translation but enforces correctness with deterministic "guardrails". Our Retrieval-Augmented Generation (RAG) pipeline generates TM Forum-compliant intents and then validates them to ensure syntactic, semantic, and policy compliance. We benchmark our solution using three open-source LLMs, demonstrating the framework’s high reliability. It achieved a 93.8% pipeline success rate with Mistral-7B, where intents were correctly processed to a definitive conclusion. The results suggest that a hybrid architecture with symbolic guardrails offers the safety and reliability necessary for deploying generative AI in network orchestration, overcoming the limitations of using isolated LLMs.
This paper presents a practical online active learning solution for adaptive channel estimation in fluid antenna systems. We model channel state information (CSI) as a spatiotemporal Gaussian process (GP) and approximate it using a deep dropout neural network (DDNN), transforming the challenge into a deep learning task. By leveraging the expressive power of neural networks, our GP model effectively captures the spatiotemporal characteristics of CSI, thereby reducing the number of ports required for accurate estimation. The DDNN, characterized by its low complexity and ease of training, employs a least-squares loss function based solely on port information and pilot measurements, eliminating the need for true CSI. This facilitates dynamic updates and allows for effective adaptation to environmental changes. Additionally, our solution incorporates a port selection method that maximizes mutual information, focusing on prediction uncertainty to optimize port usage. Simulation results demonstrate that our proposed online active learning method significantly outperforms state-of-the-art solutions in CSI estimation accuracy and exhibits robust adaptability in dynamic environments.
