Digital twins have emerged as a key enabler for monitoring, simulation, and optimization in complex systems. However, existing approaches remain fundamentally centered on state replication and prediction, lacking the ability to systematically reuse experience, interpret system behavior, and support interactive decision-making under dynamic conditions. In this article, we introduce the cognitive twin as a new system paradigm that extends state synchronization with structured cognitive capabilities, including perception, memory, causal reasoning, and interaction. Unlike traditional digital twins and their learning-based extensions, a cognitive twin maintains an evolving cognitive state that supports experience-aware reasoning, causal interpretation, and context-aware interaction, enabling a transition from data-driven monitoring toward understanding-driven operation. We develop a four-layer reference architecture that captures how cognitive processes are distributed across device, edge, and cloud domains and realized through the continuous exchange of structured state, contextual knowledge, and inferred decisions. A central insight is that this paradigm fundamentally reshapes the role of communication. Rather than solely supporting data exchange, communication directly influences the formation, synchronization, and evolution of cognitive state under heterogeneous constraints on latency, bandwidth, reliability, and synchronization. A key hypothesis explored in this article is that communication impairments affect cognitive fidelity asymmetrically with respect to semantic importance, such that degradation of semantically critical information may produce disproportionately large reasoning errors relative to equivalent losses in redundant or low-impact data streams. This perspective motivates wireless-native system designs in which communication, computation, memory, and cognition are jointly optimized rather than treated as independent subsystems. [...]
Although User-Generated Content (UGC) metaverse markets exhibit structural features that platform theory associates with network effects-driven concentration, empirical evidence on whether, and how, such concentration emerges in this context remains scarce. As a preliminary empirical investigation, this paper takes VRChat as the leading exemplar of a UGC metaverse, where avatars and compatible outfits are produced by independent creators and traded on external e-commerce, and examines how outfit creators allocate production effort across the avatar landscape over time. We propose a UGC metaverse network effects framework distinguishing direct and indirect network effects, and analyze the top 100 avatars from 2023 to 2025 using two complementary diagnostics: the Herfindahl-Hirschman Index as a creator-level static descriptor, and an avatar-level power-law regression of later on initial outfit stock as a dynamic descriptor of cumulative advantage. We document a divergence: although the static creator-level distribution remained stable despite a more than threefold market expansion, the avatar-level growth elasticity shifted significantly across year-pairs (Vy = 0.20, paired bootstrap 95% CI [0.06, 0.35], p = 0.003), rising from y = 0.81 in 2023 - 2024 (95% CI [0.66, 0.95]) to y = 1.01 in 2024 - 2025 (95% CI [0.99, 1.03]). The 2023 - 2024 estimate lies significantly below 1, consistent with a sublinear regime in which mid-tier avatars closed the gap with incumbents; the 2024-2025 estimate is essentially at the proportional-growth threshold. Within this preliminary scope, the shift is consistent with weakening avatar-level catch-up dynamics; it does not demonstrate increased creator-level concentration. [...]
The evolution toward 6G networks necessitates a shift toward fully automated, AI-native, and zero-touch operational solutions to handle unprecedented network complexity. Within the Open RAN (O-RAN) architecture, the near-Real-Time RAN Intelligent Controller (near-RT RIC) plays a pivotal role in hosting third-party xApps for radio resource management. However, the deployment of multiple xApps from different vendors often leads to operational conflicts, such as overlapping control commands and resource competition, which can degrade network performance. This paper addresses this challenge by proposing an ML-driven network orchestrator specifically designed for conflict detection and resolution in a multi-xApp 6G O-RAN environment. Our framework utilizes advanced machine learning techniques to autonomously identify potential steering conflicts and execute resolution strategies in real time. Experimental results demonstrate that the ML-driven orchestrator effectively mitigates performance degradation caused by xApp collisions, maintaining high-fidelity service delivery even under dense traffic conditions. This study contributes to the realization of self-optimizing 6G networks by providing a scalable and resilient orchestration logic for future autonomous RAN operations.
The integration of terrestrial and Non-Terrestrial Networks (NTNs) is a cornerstone of the sixth-Generation (6G) vision, yet orchestrating artificial intelligence (AI) workloads across these heterogeneous domains remains an open challenge. This paper introduces an AI-driven orchestration architecture for integrated satellite-terrestrial 6G networks that extends four components of our previously proposed AI-native design, namely the 3rd Generation Partnership Project (3GPP) enhanced Network Data Analytics Function (NWDAF), Service Hosting Environment (SHE), Network Knowledge Exposure Function (NKEF), and AI agent framework, with NTN-specific capabilities. Drawing on a systematic analysis of the 20 ubiquitous connectivity use cases and the five AI-NTN convergence use cases from the 3GPP Technical Report (TR) 22.870 document, we conceptualize these architectural extensions: (i) an NTN-enhancement to NWDAF for ingesting satellite telemetry and ephemeris data to be used in predictive handover and coverage analytics, (ii) a space-edge computing tier within SHE that enables onboard satellite AI inference with a latency-aware model placement framework, (iii) a cross-domain AI agent for intent-driven orchestration across terrestrial and satellite operator boundaries, and (iv) an NKEF feature exposing constellation topology and coverage predictions to third-party applications. In order to demonstrate the effectiveness of this orchestration architecture, we present a handover management framework addressing four transition types (satellite-to-satellite, satellite-to-terrestrial hand-out, terrestrial-to-satellite hand-in, and inter-orbit) with backhaul-aware path selection leveraging inter-satellite links. Next, a comparative evaluation of eight NTN testbed platforms identifies current validation capabilities and gaps with respect to the discussed end-to-end system. [...]
The robustness of waveform selection is crucial for reliable physical-layer design for the upcoming 6th Generation (6G) wireless technology. This study presents a comparative performance evaluation of Orthogonal Time Frequency Space (OTFS) and Orthogonal Frequency Division Multiplexing (OFDM) waveforms, using Binary Phase Shift Keying (BPSK) and Quadrature Phase Shift Keying (QPSK), in high-Doppler midband channels. The study focused on the 6G midband (7-15 GHz) spectrum, utilizing a 10 GHz carrier and 1 GHz bandwidth. The simulation results demonstrated that OTFS achieves significant performance improvements over OFDM with Bit Error Rate (BER) gains ranging from 1.14x at low Signal-to-Noise Ratio (SNR) to over 2000x at 20 dB SNR, Error Vector Magnitude (EVM) improvements of 21.4% to 68.9%, and Spectral Efficiency (SE) gains up to 7.3%. OTFS approached theoretical Additive White Gaussian Noise (AWGN) performance bounds, while OFDM saturated at BER about 6.8 x 10-2 due to Inter-Carrier Interference (ICI). The study's outcomes suggested that OTFS is a promising candidate waveform for reliable 6G wireless technology in dynamic environments, with reduced-complexity block-circulant detectors making practical deployment feasible.
The European 6G Flagship project Hexa-X-II created momentum around 6G via 44 member organizations covering the entire value chain. The project lasted from January 2023 until June 2025. Here the key findings of the project are summarized, including requirements for sustainable, inclusive and trustworthy 6G, end-to-end system view and architecture, its validation, enhanced connectivity via 6G, network sensing, compute and AI for novel digital services as well as efficient network realization, implementation and management.
5G network slicing enables flexible and efficient service deployment by maintaining logical isolation and facilitating tailored Quality of Service (QoS) and security policies. However, network slices increase network complexity and expand the attack surface, making them more susceptible to Distributed Denial-of-Service (DDoS) attacks. Due to the growing attack landscape, Machine Learning (ML)-based Intrusion Detection Systems (IDS) are becoming more popular over traditional IDS. Since network slices are used by different verticals, they often have strict privacy requirements. Conventional ML-based IDS are centralized in nature, combining all data on a central server, which fails to uphold slice-level privacy constraints. This paper proposes FLAccShield, a Federated Learning (FL)-based privacy-preserving IDS framework for DDoS detection in 5G networks through collaboration among multiple slices. The framework is rigorously evaluated for supervised and unsupervised settings for both Independent and Identically Distributed (IID) and non-IID datasets against a varied range of features. The experimental results demonstrate that FLAccShield achieves consistent and stable detection performance across varying feature set sizes and data distributions. In supervised settings with non-IID data, the most practically challenging scenario, FLAccShield outperforms FedAvg and FedProx by approximately 3% in F1 score while exhibiting substantially less performance degradation as feature count increases (7% decline versus 10-22% for baselines). In supervised IID settings, FLAccShield maintains uniform per-client performance where baselines show client-level deviation. In unsupervised anomaly detection, FLAccShield achieves comparable performance to FedAvg and FedProx (within 0.4 - 1%), while QFedAvg proves poorly suited to the autoencoder-based anomaly detection paradigm.
The transition toward zero-touch and AI-native network operations requires closing the loop between data collection, analytics, and automated control across distributed network functions. While the 3GPP Network Data Analytics Function (NWDAF) provides a standardized foundation for network intelligence, its practical realization as a network automation enabler remains largely unexplored. In this paper, we present a standards-compliant framework for closed-loop network automation that elevates NWDAF from an analytics service to an active component in runtime orchestration. Our design integrates (i) real-time user-plane observability via a novel implementation of the User Plane Function (UPF) Event Exposure Service, (ii) actionable analytics with modular engines and Machine Learning (ML) lifecycle management, and (iii) automated action and policy enforcement through control-plane functions, forming a complete data analytics action pipeline. We implement the proposed framework, integrate it with an open-source 5G core, and evaluate its end-to-end performance under realistic workloads. Our results show that fine-grained user-plane telemetry can be collected with minimal overhead while enabling responsive closed-loop control at runtime. Through two representative use cases, anomaly detection with automated mitigation and congestion-aware quality of service adaptation, we demonstrate that NWDAF-driven control can effectively translate analytics into actionable policies while preserving system scalability. This work provides the first comprehensive evidence that standards-aligned NWDAF can serve as a practical foundation for zero-touch network automation, bridging the gap between 3GPP-defined analytics and runtime orchestration.
Automated Guided Vehicles (AGVs) are widely used for material handling across various industries, such as warehouses, manufacturing plants, and automated container terminals. As AGVs are battery-powered, it is necessary to schedule AGV's tasks considering energy-related aspects. That is, understanding how an AGV's energy is consumed or recharged in a given operational environment is crucial to optimizing both operational efficiency and energy consumption. This article presents a comprehensive analysis of energy-aware AGV operations, emphasizing three critical areas: the impact of the operational environment, energy consumption modeling, and energy-aware decision-making for scheduling and path planning. This article examines the layout structures and energy supply strategies in two different environments: automated container terminals and workshops, to understand the environmental impact on AGV energy efficiency. A comprehensive review of energy consumption models provides insight into the physical characteristics of AGVs' energy consumption. Finally, we provide variations of task scheduling and path planning problems that explicitly take into account such energy aspects and introduce how literature tackles to solve those problems.
Federated learning at edge systems not only mitigates privacy concerns by keeping data localized but also leverages edge computing resources to enable real-time AI inference and decision-making. In a blockchain-based federated learning framework over edge clouds, edge servers as clients can contribute private data or computing resources to the overall training or mining task for secure model aggregation. To overcome the impractical assumption that edge servers will voluntarily join training or mining, it is crucial to design an incentive mechanism that motivates edge servers to achieve optimal training and mining outcomes. In this paper, we investigate the incentive mechanism design for a semi-asynchronous blockchain-based federated edge learning system. We model the resource pricing mechanism among edge servers and task publishers as a Stackelberg game and prove the existence and uniqueness of a Nash equilibrium in such a game. We then propose an iterative algorithm based on the Alternating Direction Method of Multipliers (ADMM) to achieve the optimal strategies for each participating edge server. Finally, our simulation results verify the convergence and efficiency of our proposed scheme.
In advanced driving assistance systems and autonomous vehicles, lane detection plays a crucial role in ensuring the safety and stability of the vehicle during driving. While deep learning-based lane detection methods can provide accurate pixel-level predictions, they can struggle to interpret lanes as a whole in the presence of interference. To address this issue, we have developed a method that includes two components: a convolutional neural network transformer and a fusion decoder. The CNN transformer extracts the overall semantics of the lanes and speeds up convergence, while the fusion decoder combines high-level semantics with low-level local features to improve accuracy and robustness. By using these two components together, our method is able to effectively detect lanes in a variety of conditions, even when interference is present. We tested our method on multiple lane datasets and obtained superior results, with the best performance on the BDD100K dataset. Our method has successfully addressed the challenge of accurately and completely detecting lanes in the presence of interference, such as darkness, shadows, and strong light. The algorithm has been employed in an edge computing device, an intelligent cart. The code has been made available at: https://github.com/squirtlecc/CNNTransformer
This paper presents AIEnergy, the first energy benchmark suite and benchmarking methodology to allow accurate energy measurement and performance evaluation of AI-empowered mobile and IoT devices with diverse AI chipsets and software stacks. We first discuss the design principles and the key challenges for developing an accurate, interpretable, and adoptable energy benchmark. We address these design challenges by developing an energy measurement methodology that incorporates three strategies and an end-user understandable scoring system. AIEnergy collects over 8.8 GB measurement data from 264 configuration combinations of eight commercial AI-empowered mobile and IoT devices with diverse chipsets, six deep learning applications with unique end-to-end processing pipelines and 12 deep neural network models under CPU, GPU, and Neural Networks API (NNAPI) delegates. AIEnergy will evolve and serve as a ready-to-adopt benchmark that is accessible by both mobile and IoT end users with non-technical backgrounds and researchers with varying levels of expertise.
Modern high-performance computing systems have undergone a significant transformation with the adoption of chiplet-based multi-die integration. This approach enables scalable improvements in computational and communication performance while reducing energy consumption and manufacturing costs. However, existing chiplet-based interconnection network designs face challenges in meeting the stringent latency and energy efficiency requirements of edge computing systems, particularly for multicast and broadcast communication. Silicon interposers impose high costs in inter-chiplet communication, primarily due to their restricted throughput capacity. Similarly, wired interconnection designs are ill-suited for managing multicast and broadcast traffic, as their design limitations, such as high hop counts, inadequate bandwidth, and resource constraints, contribute to significant power consumption and latency. Chiplet-based hybrid interconnection designs overcome these challenges by integrating both wired and wireless interconnects that can adapt to diverse communication patterns and requirements. This paper introduces a data-driven, machine learning-based dynamic routing framework that intelligently adapts to communication needs and directs traffic to wired, interposer, or hybrid communication links by analyzing workload patterns.
Augmented Reality and Mixed Reality (AR/MR) headsets are transforming computing by enabling immersive 3D experiences, yet inherent size and power limitations prevent them from matching desktop systems in delivering complex graphics. As a result, many graphics-intensive applications cannot run natively on these devices. Remote rendering offers a promising alternative by offloading heavy 3D graphics computations to a server and streaming the rendered results to AR/MR headsets. However, conventional remote rendering approaches often suffer from considerable interaction latency over wireless networks, making them unsuitable for latency-sensitive applications. This paper introduces a novel low-latency remote rendering system that enables real-time 3D graphics on AR/MR headsets. By leveraging image-based rendering with advanced 3D image warping techniques, our system synthesizes headset displays from server-generated depth images. Experimental results demonstrate that our approach significantly reduces interaction latency while maintaining high rendering quality, achieved through the careful optimization of multiple-depth image generation strategies.
Achieving high-precision ranging in Integrated Sensing and Communications (ISAC) systems operating in low-frequency bands is challenging due to fragmented frequency resources and clock synchronization errors across multiple operators. Cross-operator Carrier Aggregation (CA) offers a potential solution by combining fragmented frequency resources, but clock synchronization errors between Base Stations (BSs) cause phase discontinuities that severely degrade ranging performance. This paper proposes a novel method for cross-operator CA scenarios that leverages successive echo signals to mitigate these phase discontinuities without requiring perfect clock synchronization between BSs. A comprehensive mathematical analysis of phase discontinuities in multi-BS systems is provided and their impact on range profiles is quantified, revealing a characteristic "fishbone effect" that substantially reduces ranging accuracy. Our proposed method effectively eliminates this effect by progressively matching phase transitions between signals from different BSs. Through extensive simulations in both single-target and multi-target scenarios, the proposed method achieves similar range accuracy comparable to systems under perfect synchronization conditions, offering a viable solution for applications requiring accurate positioning in resource-constrained environments.
Autonomous Vehicles (AVs) are revolutionizing transportation, but their reliance on interconnected cyber-physical systems exposes them to unprecedented cybersecurity risks. This study addresses the critical challenge of detecting real-time cyber intrusions in self-driving vehicles by leveraging a dataset from the Udacity self-driving car project. We simulate four high-impact attack vectors, Denial of Service (DoS), spoofing, replay, and fuzzy attacks, by injecting noise into spatial features (e.g., bounding box coordinates) to replicate adversarial scenarios. We develop and evaluate two lightweight neural network architectures (NN-1 and NN-2) alongside a logistic regression baseline (LG-1) for intrusion detection. The models achieve exceptional performance, with NN-2 attaining an AUC score of 93.15% and 93.15% accuracy, demonstrating their suitability for edge deployment in AV environments. Through explainable AI techniques, we uncover unique forensic fingerprints of each attack type, such as spatial corruption in fuzzy attacks and temporal anomalies in replay attacks, offering actionable insights for feature engineering and proactive defense. Visual analytics, including confusion matrices, ROC curves, and feature importance plots, validate the models' robustness and interpretability. This research sets a new benchmark for AV cybersecurity, delivering a scalable, field-ready toolkit for Original Equipment Manufacturers (OEMs) and policymakers. By aligning intrusion fingerprints with SAE J3061 automotive security standards, we provide a pathway for integrating machine learning into safety-critical AV systems. Our findings underscore the urgent need for security-by-design AI, ensuring that AVs not only drive autonomously but also defend autonomously. This work bridges the gap between theoretical cybersecurity and life-preserving engineering, offering a leap toward safer, more secure autonomous transportation.
The privacy and security concerns related to execution data, model parameters, and processing algorithms have assumed an increasing pivotal role in the rapid development and massive deployment of Large Language Model (LLM) technology. This paper provides an overview of private LLM by addressing its working principles, application scenarios, security requirements, training and inference algorithms, and more importantly, its silicon and chip implementation in order to bring this game-changing technology to real-world products. This paper is organized into five sections. The first section is devoted to the background and motivations for proposing private LLM technology. According to different requirements for privacy and security, in the second section we categorize the proposed private LLM technology into three application scenarios (security levels) and then present the corresponding algorithms related to training and inferences. In order to converge private LLM into optimum silicon implementation, we present the proposed Cornami solution and its comparisons with existing solutions (GPU and ASIC) in terms of power consumption, cost and processing latency in Section 3 and Section 4. In the last section, we make some conclusions and note further discussions.
With the emergence of 6G, there is an increasing need for autonomous network service quality assurance. This process is particularly challenging at scale when: (i) requirements of tens of industrial components are to be considered; (ii) conflicting requirements are to be resolved; (iii) mapping needs to be done across industrial and 6G domains; and (iv) autonomous intent resolution is to be incorporated. In this paper, we propose STRAUSS, a scalable intent-driven framework for service quality assurance. We exploit the features of asset administration shells to effectively manage requirements towards the network in an interoperable manner. The intents are mapped effectively between domains using AI-driven techniques. A similar Artificial Intelligence (AI)-driven technique is used to decompose intents with end-to-end expectations to various resource level domains to enable autonomous network configuration management. The proposed system is demonstrated over a realistic use case with multiple picking robots and Autonomous Mobile Robots (AMRs).
Vision Transformers (ViTs) have evolved in the field of computer vision by transitioning traditional Convolutional Neural Networks (CNNs) into attention-based architectures. This architecture processes input images as sequences of patches. ViTs achieve enhanced performance in many tasks such as image classification and object detection due to their ability to capture global dependencies within input data. While their software implementations are widely adopted, deploying ViTs on hardware introduces several challenges. These include fault tolerance in the presence of hardware failures, real-time reliability, and high computational requirements. Permanent faults that are in processing elements, interconnections, or memory subsystems lead to incorrect computations and degrading system performance. This paper proposes a fault-tolerant hardware implementation of ViTs to overcome these challenges. This hardware implementation integrates real-time fault detection and recovery mechanisms. The architecture includes four primary units: patch embedding, encoder, decoder, and Multi Layer Perceptron (MLP) which are supported by fault-tolerant components such as lightweight recompute units, a centralized Built-In Self-Test (BIST), and a learning-based decision-making system using machine learning model 'decision tree'. These units are interconnected through a centralized global buffer for efficient data transfer, ensuring seamless operation even under fault conditions.
For machine learning with tabular data, a table transformer (TabTransformer) is a state-of-the-art neural network model, while Differential Privacy (DP) is an essential component to ensure data privacy. In this paper, we explore the benefits of combining these two aspects together in the scenario of transfer learning, differentially private pretraining and fine-tuning of TabTransformers with a variety of Parameter-Efficient Fine-Tuning (PEFT) methods, including adapter, LoRA, and prompt tuning. Our extensive experiments on four ACS datasets with different configurations show that these PEFT methods outperform traditional approaches in terms of the accuracy of the downstream task and the number of trainable parameters, thus achieving an improved trade-off among parameter efficiency, privacy, and accuracy.
