The shift toward 6 G architectures demands a transition from human-centric management to AI-native, zero-touch orchestration to navigate extreme network complexity. While the Open RAN (O-RAN) framework utilizes the near-real-time RAN intelligent controller (RIC) to host third-party xApps for resource optimization, the lack of coordination between multivendor applications often triggers operational collisions and resource competition. This paper introduces a novel ML-driven network orchestrator designed as an intelligent interceptor layer that decouples xApps from the RIC. The proposed framework transforms asynchronous xApp requests into a collective optimization problem, evaluating the combinatorial impact of control actions before execution. The system leverages high-speed eXtreme Gradient Boosting (XGBoost) predictors and an inmemory feature store to forecast network utility within a 5 -second prediction horizon, ensuring stability through a dedicated cooling mechanism for rejected requests. Experimental evaluations conducted via ns-3 and FlexRIC integration to demonstrate that the orchestrator effectively prevents performance degradation, yielding an increase in mobility stability and an additional gain in overall throughput. Our results validate that this scalable orchestration logic provides a resilient foundation for autonomous 6G operations by prioritizing global system utility over individual xApp objectives while effectively mitigating xApp conflicts and preventing resource contention.
Trusted xApp distribution is a prerequisite for O-RAN adoption, but current proposals rely on logically centralized stores, a single point of trust unsuitable for multi-operator deployments. We propose a decentralized xApp marketplace governed by Ethereum smart contracts, where a consortium of operators jointly controls xApp onboarding, validation, licensing, and revocation with no central authority; artifact hashes are stored on-chain while binaries reside on IPFS. Three smart contracts are deployed on Sepolia testnet and evaluated for gas cost across all lifecycle operations and for resilience against six threat classes (Sybil attacks, artifact injection, governance takeover, replay, key compromise, and admin role abuse). Post-approval deployment latency is statistically indistinguishable from a centralized baseline $(\Delta=-1.8 \text{ms})$, while the on-chain audit trail provides immutable provenance that centralized stores cannot offer.
In this paper, we introduce a dynamic approach to routing in SDNs and VNF placement, using a parameterized multi-objective optimization model, which can handle different weights for objectives such as energy, delay, congestion, and resource constraints. This approach allows the network to behave dynamically in response to changing conditions, since the optimization and objective weighting are decoupled processes. The online control layer, which uses rule-based approaches and reinforcement learning, can dynamically control the objectives in real time and improve decision-making efficiency. The results demonstrate the effectiveness of the proposed approach in terms of the costs of service deployment and overall network performance.
This paper explores the application of Digital Twins in 6 G wireless communications. Simulating these environments is crucial for performance evaluation and sensing applications. We introduce Raycom, a Command Line Interface (CLI) tool based on the Sionna RayTracing engine, designed to automate wireless channel evaluation and generate datasets. Utilizing Raycom, we tackle a complex inverse problem: reconstructing a visual representation of an indoor space entirely from Radio Frequency ($\mathbf{R F}$) channel state data, enabling privacypreserving spatial awareness. By employing a Conditional Generative Adversarial Network (Pix2Pix), we map lowresolution RF heatmaps to high-resolution ($256 \times 256$) optical images. The network converges over 200 epochs on 400 spatial samples. We present a limitation analysis on complex geometries across multiple viewpoints. Our results highlight the application potential and the informational bottlenecks imposed by small antenna arrays ($5 \times 5$), laying the groundwork for high-fidelity wireless vision using Massive MIMO systems.
Recent advances in communication technologies and networked intelligent systems are enabling new paradigms for connected mobility, where distributed sensing and real-time data processing play a critical role. In such environments, efficient data transmission and low-latency operation are essential, particularly for UAV-assisted applications that rely on continuous environmental monitoring. To address these requirements, perception is increasingly performed at the edge, allowing UAV platforms to operate as intelligent sensing nodes that process visual data locally and transmit only relevant semantic information. Within this context, fisheye cameras provide wide-area coverage suitable for mobility monitoring, while introducing additional challenges due to geometric distortion. This work presents a comparative evaluation of fisheye-based object detection methods under realistic deployment conditions on an embedded platform (NVIDIA Jetson Orin Nano), analyzing distortion-aware and foundation-model-based approaches and highlighting accuracy-efficiency trade-offs for scalable and communication-efficient intelligent transportation systems.
This paper presents a comprehensive experimental evaluation of RIS sharing methodologies in a realistic wireless testbed environment. We specifically assess and benchmark three distinct RIS resource-sharing strategies: Segmentation, Time Division Multiple Access, and codebook multiplexing algorithms, that have been proposed recently in the research bibliography. Our evaluation addresses scenarios involving multiple users with both identical and varying pricing tiers. Measurements obtained from a real-world RIS setup demonstrate that multiplexing algorithms effectively achieve fairness and proportional resource allocation across users, outperforming segmentation and TDMA methods. These experimental results validate multiplexing as a promising and equitable solution for integrating RIS units into future 6G networks, addressing latency constraints and varying stakeholderdefined pricing policies.
We address the near-far clutter masking problem in bistatic V2V Integrated Sensing and Communication (ISAC) over 5G NR sidelink: strong clutter echoes (large vehicles, guardrails, direct path) bury weak targets such as pedestrians or cyclists in sidelobe leakage that receiver-side cancellation cannot recover. We propose a transmit-side range-domain null-steering solution that exploits the multi-port DM-RS structure of NR sidelink: per-port unimodular phase weights at the transmitting vehicle place a null at the dominant clutter delay in the bistatic range profile observed at the receiving vehicle, with no protocol modification. Under explicitly stated synchronization and phase-coherence assumptions, the reconfiguration is transparent to the receiver's channel estimator. The resulting Unimodular Quadratic Program on $\mathbb{T}^{4}$ admits a tight Semidefinite Relaxation, enabling per-slot phase optimization in $<50 \mu$ s. NR-compliant simulation confirms 29.9 dB clutterregion suppression with 1.9 dB sensing gain variation at the target of interest, and BER indistinguishable from the QPSK reference.
We introduce the Wireless Ad hoc Network Digital twins Architecture (WANDA), a Digital Twins (DT) architecture specifically tailored to the unique aspects of wireless ad hoc networks, in particular the facts that their data transporting capabilities are hard to describe and change fast, and that data rates are relatively low, compared to other networks. WANDA consists of three layers, the Physical Twin (where the wireless nodes are placed), the Digital Twin (comprising the state, a controller, a set of agents, a publish/subscribe broker, and a dashboard), and an application layer where a number of users reside that use the DT as a service. We design and implement a prototype of the architecture using Raspberry Pis as the nodes of the network, and conduct a set of experiments to validate its performance in an experimental setup. In particular, we demonstrate how two instances of an epsilon-greedy Multi-Armed Bandit (MAB) agent can autonomously optimize data transmission rates by interacting with network nodes, in two respective tasks: selecting the channel that maximizes a transmitter-receiver long-term throughput, and selecting the relay that maximizes a long-term source-destination end-to-end throughput.
Concurrent xApp operation on the Near-Real Time RAN Intelligent Controller (Near-RT RIC) in Open Radio Access Networks (O-RAN) frequently produces conflicting control decisions that degrade network performance and user experience. We propose MACOL-x, a lightweight decentralised multi-agent coordination framework for Open RAN xApp conflict resolution. Each xApp independently encodes the recent actions of its peers as a coordination context and learns-without any central orchestrator-which action-context combinations are conflictprone. A hybrid two-tier context memory provides $O(1)$ exact lookup for warm contexts and $K$-Nearest Neighbour interpolation for cold starts, feeding a conflict-targeted backoff gate that suppresses conflict-creating actions while leaving conflict-neutral ones intact. Evaluations against MAB, Q-Learning, MACOL, and QMIX baselines in a heterogeneous O-RAN HetNet show that MACOL-x achieves 95% conflict reduction, 91% fewer handovers, and $11 \times$ lower user-plane interruption time, with reward comparable to QMIX (+89.5 vs. +89.6) while generating $6 \times$ fewer conflicts and operating at $10 \times$ lower per-step computation (4.56 ms vs. 45.99 ms) -without any neural network, replay buffer, or central coordinator.
As industrial systems increasingly integrate Information Technology (IT) and Operational Technology (OT), traditional identity management approaches face significant challenges in securing complex, heterogeneous environments. These systems require robust authentication mechanisms capable of managing multiple identities across diverse devices, applications, and networks while maintaining operational continuity. This paper presents a comprehensive analysis of blockchainbased Self-Sovereign Identity (SSI) integration in industrial systems, addressing critical security gaps through decentralized, cryptographically-secured identity management. We examine four key use cases: cell/area zone segmentation, administrative user access control, infrastructure service authentication, and plant-wide application security. Our experimental evaluation demonstrates that blockchain-based SSI reduces authentication latency by up to 95% under denial-of-service attacks compared to traditional systems, while providing immutable audit trails and eliminating single points of failure. The proposed approach offers a scalable, tamper-proof solution that enhances industrial cybersecurity without compromising operational requirements.
This paper addresses the spectrum coexistence challenge between 5G NR sidelink integrated sensing and communication (ISAC) systems and legacy frequency-modulated continuous-wave (FMCW) automotive radars in the 24 GHz band. We first show a domain mismatch between existing spatial-and channel-domain interference metrics and the delay-Doppler matched-filter output of legacy FMCW radars. This motivates the cross-ambiguity function (CAF) Gramian as a metric directly aligned with FMCW radar processing. A non-convex symbol-level precoding framework is then designed for practical hybrid analog-digital architectures, jointly addressing CAF-based integrated sidelobe level (ISL) minimization and constant-modulus (CM) phase-shifter constraints. The resulting problem is solved using a projected gradient descent algorithm with alternating feasibility projections. Simulation results demonstrate that the proposed CAF-aware design significantly improves radar detection performance compared to spatial and channel-domain baselines, while preserving communication quality of service under practical hardware constraints.
Data-driven research in mobile networks depends critically on access to high-quality Performance Management (PM) data. However, regulatory, privacy, and confidentiality constraints limit data sharing, hindering reproducibility and systematic comparison across studies. Synthetic Data Generation (SDG) offers a promising alternative, yet publicly available datasets and systematic evaluation frameworks for mobile network PM data remain lacking. This paper introduces Cellular Operational Observations for Performance and Evaluation Research (COOPER), an open-source synthetic dataset of multi-cell 5 G PM time series comprising 45 indicators across multiple PM categories. COOPER is designed to reproduce the statistical, temporal, and inter-variable characteristics of an operational network. Three SDG paradigms, adversarial, probabilistic, and model-based, are evaluated using a unified framework assessing distributional similarity, temporal fidelity, dependency preservation, protocol consistency, and forecasting accuracy. Results show that the selected approach consistently outperforms alternatives while satisfying protocol constraints. The COOPER dataset is now publicly available on Hugging Face 11https://huggingface.co/datasets/CelfAI/COOPER
Adaptive, uncertainty-aware, and certificationcompatible validation is central in safety-critical UAS communication software. We observe existing test prioritisation techniques do not fully meet these requirements collectively. This study presents a Constrained Bayesian contextual Bandit framework that combines Bayesian logistic regression, risk-sensitive upper confidence bound selection and budget constraints. Continuous improvement without retraining is made possible by a cross-session permanent posterior and an auditable foundation for fault impact weighing is provided by a dynamic severity function that combines requirement criticality and code commit information. Structured, requirement-traced artefacts that are compatible with formal safety cases are produced via a dedicated evidence layer. The framework is expected to demonstrate enhanced fault detection efficiency than current techniques, in sparse failure regimes, alongside measurable cost efficiency gains from hard constraint enforcement. This constitutes a principled foundation for certifiable adaptive testing of safety-critical UAS systems.
Evaluating Intrusion Detection and Prevention Systems (IDS/IPS) in Operational Technology (OT) environments requires domain-aware threat models, reproducible testbeds and rigorous ground-truth annotation, capabilities absent from mainstream Information Technology (IT)-focused benchmarking datasets. This paper proposes a generic IDS/IPS benchmarking framework applicable across IoT, Industrial Control Systems (ICS) and industrial OT deployments. It demonstrates its applicability through the ELEVATE project: a cybersecurity evaluation initiative targeting an IoT-connected elevator ecosystem. The framework defines a structured four-layer pipeline encompassing risk-driven scenario design, reproducible infrastructure, automated execution with integrated STIX 2.1 ground-truth generation and a pluggable evaluation engine that can accommodate any detection system. Within ELEVATE, the framework is instantiated with five attack scenarios derived from an IEC 62443-3-3 risk assessment. The Pragma IoT AI IDS, a machine-learning flow-level classifier, is used as a representative detection system to validate the framework. Results show perfect precision (1.0) on high-rate web attacks, F1-scores up to 1.0 for brute-force scenarios and F1 up to 0.96 for Mirai malware variants, while revealing a generalisation gap on held-out samples that motivates future work on continual learning. The framework and its instantiation methodology are directly transferable to any IoT or ICS domain requiring structured, evidence-based security evaluation.
Tiny Machine Learning (TinyML) brings AI capabilities to low-power edge devices, enabling local data processing without cloud reliance. This paper evaluates a distributed Recurrent Trend Predictive Neural Network (rTPNN) deployed on ESP8266 and ESP32 microcontrollers within a smart home IoT network. The rTPNN achieves superior accuracy with fewer parameters than standard models while incurring negligible computational costs. Energy profiling demonstrates that constant WiFi connectivity extends battery life eight-fold compared to periodic reconnection. However, security assessments reveal that constant connectivity is highly susceptible to Denial of Service (DoS) saturation, and parameter injection can degrade accuracy by over 70%. These results substantiate the feasibility of rTPNN as a TinyML for sustained operation while highlighting the critical need for codesigned security mechanisms.
The ELEGANT project (Evolution of Infrastructure and Expanded Mobile System Toward Sixth Generation Artificial Intelligence Networks) develops a unique 6 G scientific infrastructure with ultra-low latency. Deployed across two campuses of the Universitat Politècnica de Catalunya (UPC), it integrates a 5G Stand-Alone radio network with millimeter-wave capabilities, a Dense Wavelength Division Multiplexing (DWDM) optical transport network, and data centers optimized for AI and edge computing, enabling advanced use cases such as connected and autonomous mobility. Operating as “6G LABNET UPC”, the platform offers R&D services including network slicing, beyond5G/6G certification, and AI model validation. As a primary use case, the SNS-JU project 6G-EWOC leverages this infrastructure for its connected and sensing mobility demonstrations.
Coherent transfer-matrix solutions for subterahertz penetration loss exhibit Fabry-Pérot ripples that are fragile under millimeter-scale construction tolerances. The incoherent energy envelope eliminates these ripples, yet the conditions under which this approximation is valid have not been rigorously quantified for the 100-300 GHz regime. This paper derives closed-form validity bounds showing the additive envelope error is below 0.04 dB for all ITU-R P.2040-4 building materials at sub-THz frequencies, and shows that Gaussian thickness jitter exceeding 0.3-0.8 mm suffices for full phase decoherence. We extend the framework to oblique incidence with polarization-resolved (TE/TM) Fresnel coefficients, revealing Brewster-angle asymmetry exceeding 4 dB. The incoherent intensity-matrix cascade is applied to multilayer partitions and validated against Monte Carlo simulations. Comparisons with 3GPP TR 38.901 and published sub-THz measurements confirm the physical consistency of the proposed framework.
Non-binary coded modulation can significantly improve the performance of orthogonal modulations, however, the high complexity of non-binary decoders remains a major obstacle for practical deployment. In non-binary polar codes, this difficulty mainly arises from the quadratic complexity of check-node (CN) processing in successive-cancellation decoding. This paper introduces Input-Truth Polarization-Aware (IT-PA) pruning, a cluster-level offline pruning method that records the bubble containing the true input pair rather than bubbles frequently selected by truncated CN processing. The resulting contribution matrices are sparse and yield compact decoding masks. For a (64,42) NB polar code over GF(64) with Cyclic Code Shift Keying (CCSK) modulation, IT-PA preserves the Frame Error Rate (FER) performance of the Extended Min-Sum (EMS) decoding and matches the Fast Successive Cancellation-Polarization Aware (FSC-PA) while reducing the number of active CN bubbles from 4944 to 2130, i.e., a $2.32 \times$ reduction. Additional bubble-count evaluations on (64,16) over $G F(64)$ and (64,48) over $G F(256)$ show that the same sparsity trend persists across the other tested configurations. For the (64,42) reference case, this corresponds to about 5.5 CN bubble evaluations per coded bit.
Among the candidate spectrum bands for sixthgeneration (6 G) networks, millimeter-wave (mmWave) frequencies are widely considered for high-capacity links, yet exhaustive beam sweeping introduces significant overhead in reconfigurable intelligent surface (RIS)-assisted systems. This paper proposes a lightweight machine learning (ML) regression framework that directly maps received power measurements from a small set of sector-aligned probing beams to the user equipment (UE) elevation angle, enabling RIS phase optimization without explicit channel state information (CSI). Four classical regressors-Knearest neighbors (KNN), decision tree (DT), random forest (RF), and support vector regressor (SVR)-are evaluated under an angle-stratified protocol across different beam configurations. KNN consistently achieves the best performance, attaining a mean absolute error (MAE) of 0.79° and $R^{2}=0.99$ with six probing beams, while reducing beam-probing overhead by over 93 % compared to exhaustive search. The predicted angles are used to construct 1-bit RIS phase matrices, yielding radiation patterns within $1^{\circ}-2^{\circ}$ of the ground-truth direction, confirming that accurate angular localization translates directly into effective beamforming.
We present a Closed Loop Intent Based Agent for autonomous resilience in 5G Core networks. By leveraging the Uplink Classifier (ULCL) and a custom Dynamic Steering API, our system detects “gray failure” on the N9 interface and triggers real-time path relocation. As gray failure we define a subtle performance degradation, such as high packet loss, occurring while a network link remains technically active and seemingly operational. Experimental results using free5GC [1] demonstrate that our agent can restore a failing video service from a Mean Opinion Score (MOS) of 2.0 to 3.5 in under 10 seconds, proving the viability of self-healing softwarized user planes.
