Accurately and continuously assessing cyber risk remains a core challenge in modern network defense. Conventional Cyber Threat Intelligence (CTI) products rely heavily on observed Indicators of Compromise (IoCs), which provide only partial or lagging visibility into malicious activity. As a result, defenders often lack persistent, data-driven measures of how threat exposure evolves across network entities such as Autonomous Systems (ASes). This paper introduces a probabilistic framework for estimating the security risk of ASes using time-varying latent risk scores derived from network threat intelligence. The method integrates heterogeneous threat signals, including IP reputation measures, observed threat events, and associated severity indicators, within a Bayesian Gaussian Random Walk model that infers a latent risk trajectory for each AS. Evaluation using both synthetic and real-world threat datasets demonstrates that the framework captures meaningful temporal trends, quantifies uncertainty during periods of sparse observation, and provides a continuous view of AS-level threat exposure.
Information-Centric Networking (ICN) has attracted significant attention as a network architecture optimized for efficient and scalable content dissemination. Moving beyond the conventional paradigm of name-based content retrieval, ICN can also support communication patterns associated with services and in-network computation. This paper aims to strengthen ICN by leveraging its core principles and introducing complementary mechanisms to improve efficiency, reliability, and performance within IP-based infrastructures. We present ICNx, an extension of the baseline ICN architecture that incorporates functional capabilities supporting adaptive and coordinated communication for both content delivery and function invocation. The study includes the architectural design and a prototype implementation on the open-source Cefore platform. In addition, a large-scale research initiative under Japan’s JST Moonshot R&D program is discussed as an illustrative use case motivating the proposed design. The results demonstrate the practical feasibility of the approach and highlight the evolving role of ICN, not merely as a content-delivery mechanism, but as a foundational communication substrate for scalable and resilient applications.
In this paper, we propose a performance metric in order to evaluate the performance of deadline-aware scheduling methods.By using linear programming to find the ideal off-line scheduling and evaluating the performance of real-time scheduling methods based on ideal scheduling, it is possible to clearly identify the differences from the ideal scheduling.This paper proposes a performance metric that focuses on four evaluation metrics and formulates liner programming to determine optimal off-line scheduling for a baseline for evaluating online scheduling algorithm. It also examines optimal scheduling for both known and unknown job occurrence patterns. Additionally, we report the performance evaluation of EDF algorithm, a representative deadline-aware scheduling method, using this metric.
This work provides the first unified comparison of DQN, DDPG, PPO, and SAC for 5G NR uplink and downlink modulation control, revealing how core RL design choices uniquely impact physical-layer Quality of Service (QoS). Deep Reinforcement Learning (DRL) algorithms optimize QoS by dynamically switching modulations in 5G New Radio (NR) systems. Evaluations span both Physical Downlink Shared Channel (PDSCH) and Uplink Shared Channel (PUSCH). Metrics include Bit Error Rate (BER), throughput, latency, and path loss. Results show DDPG consistently outperforms others, offering insights into the role of policy entropy and exploration in wireless control systems.
This paper presents a novel framework for generating realistic video footage at novel viewpoints from 3D Gaussian splats using text prompts. Current approaches face two key challenges: incorporating high-dimensional semantic information into dense Gaussian representations without costly annotations and efficiently passing representations to a language model to generate smooth, aesthetically pleasing camera trajectories. To address these, we propose a two-component pipeline. First, our Sparse Spatial-Semantic (SSS) attribute Augmentor (SAug) enriches Gaussian splats with spatial-semantic features learned from multi-view images alone, along with our sparsity control technique maintaining compact representations despite millions of splats. Second, our LLM-based Motion Generator (L-MR) leverages Large Language Models to produce smooth 6D camera pose trajectories directly from text prompts and compact object-based representations using SAug outputs. Together, these components enable text-driven generation of realistic novel-view videos, with SAug providing semantically-aware scene representations and L-MR translating natural language into cinematic camera motions. Experiments demonstrate that our method generates semantically and spatially more accurate trajectories while keeping the memory and computation footprint of the rendering pipeline low.
In this study, we focus on classifying the parameters of symbols transmitted over Orthogonal Frequency Division Multiplexing (OFDM) (Cyclic Prefix-OFDM (CP-OFDM) and Discrete Fourier Transform-OFDM (DFT-OFDM)) subcarriers, specifically adhering to 5G uplink specifications. To achieve this, 5G in-phase/quadrature (I/Q) data and measured Signal-to-Noise Ratio (SNR) are input into a convolutional neural network (CNN)-based classifier. Notably, this classifier operates without the need for prior knowledge of protocol-specific information, such as the resource allocation details of $5 G$ physical channels. Our approach has been evaluated using both synthetic and over-the-air datasets. For the over-the-air CP-OFDM I/Q samples, the classifier achieves a minimum accuracy of approximately 0.7 at an SNR of 5 dB, which is higher than the reported state-of-the-art results under similar conditions.
Recent advances in display hardware have significantly outpaced improvements in content delivery, particularly in bandwidth-constrained streaming environments. As a result, modern display systems increasingly rely on AI-based enhancement techniques such as super-resolution, motion interpolation, and reconstruction to bridge the gap between content quality and display capability. While existing quality frameworks - most notably High Dynamic Range (HDR)-successfully describe luminance, contrast, and color volume, they do not adequately capture whether reconstructed visual details remain perceptually stable, structurally coherent, or credible to human observers. In this paper, we introduce Rich Detail Range (RDR) as a descriptive perceptual concept for discussing picture quality in AI-enhanced display pipelines. RDR characterizes the perceived richness, integrity, and fidelity of visual details across spatial, temporal, and structural dimensions, with an emphasis on content-aware processing. RDR is not proposed as a standard, algorithm, or evaluation metric, but rather as a conceptual framework intended to complement existing quality dimensions and support clearer discussion of emerging perceptual challenges in modern display systems.
Push-based multi-factor authentication (MFA) is vulnerable to habituation and “push-fatigue”: when notifications arrive, users approve reflexively. We ask how the mobile context—small screens and app switching—affects different user segments in compare-and-confirm (number matching) under a first-factor-compromise threat model. Using a controlled withinsubjects study ($\mathrm{N}=65; 24$ trials/participant; benign vs. attack; PC vs. phone), we analyze attack-trial correctness and decision time. The most surprising result is not a uniform mobile penalty but a mobile improvement for medium-skill participants: on attack trials they are almost 9 percentage points more likely to respond correctly on phones than on PCs, while high-skill users show little device difference and the small low-skill group exhibits ceiling effects. We hypothesize a Forced Focus mechanism: on phones, transient, full-screen prompts and tighter interaction loops reduce competing stimuli and force attention onto the compare-and-confirm task, offsetting switch costs for some users. Time-to-decision signatures and qualitative comments about app switching and small-screen comparison are consistent with this account. We translate these findings into deployable mitigations that treat interface friction as a design parameter: unify the comparison surface (deep-link/overlay), surface lightweight origin/time context, and apply device- and segment-aware escalation where risk signals or segments warrant it—hardening mobile MFA without broadly slowing everyone down.
This paper addresses the general security problem in mobile edge computing. The problem is modeled as a two-stage attacking-defending Stackelberg security game, and a multiagent reinforcement learning framework based on independent proximal policy optimization is designed to solve the problem. The framework employs Markov decision processes for both the attacker and the defender to systematically represent network states, attack vectors, and resource allocation actions. The reward functions are designed to reflect the objectives of network disruption for the attacker and network protection for the defender. Experimental results demonstrate that our proposed framework achieves stable convergence and significantly reduces attack success rates compared to baseline methods in the randomly generated edge computing network.
The design of network architectures and standard communication protocols follow requirements and specifications from the International Telecommunication Union (ITU) to establish the corresponding Quality of Service (QoS) requirements and ensure the interoperability and continuity of services. Availability being one of the ITU QoS metrics, it is revisited in this paper to explain the need to clarify the 6 G architecture design principles such as centralized or distributed approaches, the communication concept as point-to-point of the physical layer and end-to-end of the network and transport layers and their effect on the expected performance in 6 G networks, particularly in the emerging 3 CN software and programmable network approach. The SimpleRAN architecture is proposed to showcase the need to distribute the core network to meet or closely approach the required QoS of the future complex 6 G architecture.
This paper introduces a novel autonomous cycle counting method for warehouse inventory management, specifically designed for double-deep and bulk storage configurations. The proposed methodology processes 2D RGB images captured by autonomous robots navigating warehouse aisles to accurately count pallets within bin locations. Our two-stage approach first performs pallet detection using Faster R-CNN and YOLO models enhanced for low-light conditions, then applies configurationspecific counting algorithms. For bulk storage, we developed a dimension-based KNN classifier that exploits the inverse relationship between pallet depth and apparent size. For double-deep storage, we implemented a color-based KNN classifier leveraging RGB attenuation with depth. Faster R-CNN achieved 97.2% average precision, outperforming both standard YOLO (83.8%) and YOLO with dark enhancement (90.8%). The counting algorithms achieved 86% accuracy for bulk storage and 100% accuracy for optimized double-deep storage. Our experimental results demonstrate that combining enhanced detection models with configuration-specific counting algorithms provides an effective solution for autonomous warehouse inventory management, though challenges remain in handling imbalanced training data and generalizing to diverse warehouse environments.
Cyberbullying affects $20-40 \%$ of digitally connected adolescents, yet existing automated detection systems achieve highly variable accuracy and lack ethical decision-making mechanisms. This study presents a unified AI framework combining Machine Learning (ML), Deep Learning (DL), and hybrid models to address cyberbullying detection across four major platforms: Wikipedia Talk, Twitter, Facebook, and YouTube. The framework introduces a novel confidence-based moderation layer that converts prediction certainty into three actionable tiers: automatic blocking (high confidence), user warnings (moderate confidence), and content publishing (low confidence). By addressing persistent challenges in platform adaptability, model interpretability, and ethical decision-making, the proposed framework provides a scalable and robust solution for real-time cyberbullying detection that strikes a balance between high technical performance and responsible, real-world deployment.
After COVID-19, remote education has emerged as new pedagogical paradigm. However, traditional teacher-to-many-students instructional systems remain ill-equipped to deliver personalized guidance and granular feedback, thus online teaching quality needs to be enhanced urgently. To overcome this limitation, we propose an AIpowered personalized remote education framework that leverages artificial intelligence’s generalization capabilities and adaptability to achieve large-scale personalized instruction. Within our proposed twostage teacher-AI-student paradigm, the teacher first instructs a general AI agent, which is subsequently specialized to tutor individual students. We validate this framework through a personalized remote piano instruction case: A Transformer-based music transcription model converts performance audio into symbolic MIDI sequences, while the teacherinstructed AI agent refines these sequences into expressive performance versions. The resulting teacherstudent performance comparison data provides students with personalized feedback on note onset timing and dynamics. To assess the effectiveness of the proposed framework, we collected piano proficiency examination data encompassing approximately 30 distinct pieces by different composers. Each piece’s performance is treated as an independent student, thereby allowing us to analyze and evaluate the effectiveness of our proposed framework in providing personalized feedback to students. Experimental results across multiple piano pieces demonstrate that this framework effectively captures performance variations, delivers explainable feedback, and balances efficiency with personalized needs in remote learning environments.
Although our lives are increasingly transitioning into the digital world, many digital assets still relate to objects or places in the physical world, e.g., websites of stores or restaurants, digital documents claiming property ownership, or digital identifiers encoded in QR codes for mobile payments in shops. Currently, users cannot securely associate digital assets with their related physical space, leading to problems such as fake brand stores, property fraud, and mobile payment scams. In many cases, the necessary information to protect digital assets exists, e.g., via contractual relationships and cadaster entries, but there is currently no uniform way of retrieving and verifying these documents. In this work, we propose the Geo-Enabled Cryptographic Key Oracle (GECKO), a geographical PKI that provides a global view of digital assets based on their geolocation and occupied space. GECKO allows for the bidirectional translation of trust between the physical and digital world. Users can verify which assets are supposed to exist at their location, as well as verify which physical space is claimed by a digital entity. GECKO supplements current PKI systems and can be used in addition to current systems when its properties are of value.
The automation of game-playing agents is a key domain for advancing AI algorithms. This research applies Convolutional Neural Networks (CNNs) to create an agent for the Chrome Dino game, overcoming the limitations of traditional rule-based systems. We address the core challenge of processing real-time visual data to execute precise jump and duck actions in a dynamic environment. Our solution is an end-toend CNN model that learns control policies directly from pixels, eliminating the need for hand-crafted features or internal game access. Experimental results confirm the model’s effectiveness, with performance metrics demonstrating proficient real-time decision-making and strong potential for visual-based automation in rapidly changing environments.
This paper presents a binary offloading strategy within the binary spatial allocation framework (BSAF) to enhance energy efficiency and task accuracy in mixed reality (MR) applications. We utilize the Microsoft HoloLens 2 and an edge server configured with You Only Look Once (YOLO) models. Our proposed framework dynamically selects local or edge execution based on scene complexity and resource constraints. The decision is driven by a closed-form utility function that incorporates accuracy, latency, and energy consumption, with Lagrangian relaxation ensuring constraint feasibility. Experimental results demonstrate that binary offloading achieves competitive accuracy ($89-91 \%$) compared to full offloading ($92-94 \%$), while reducing average latency by up to 37% and energy consumption by 44%, respectively. Furthermore, binary offloading sustains a 48% battery level after 50 minutes while the battery is completely depleted under full offloading.
In this research, we introduced FedAdaptCAD, a novel Federated Learning (FL) framework designed for Anomaly Detection (AD) in critical infrastructure systems. This framework emphasizes robustness, scalability, and privacy preservation in mitigating anomalies. Key innovations include Client-Aware Aggregation (CAA), which adjusts client contributions based on data quality and local model performance, and Dynamic Model Adaptation (DMA), which optimizes global model hyperparameters in real-time to respond to changing attack patterns. Evaluated on three benchmark datasets: WADI, CMAPSS, and BATADAL, FedAdapt-CAD demonstrates superior global accuracy and fairness among heterogeneous clients. The experimental results indicate a significant improvement in detection rates for rare and complex attacks, achieving up to a 5% enhancement in precision and recall compared to traditional FL methods such as FedAvg, FedProx, and FedH2L.
We present a Hierarchical Actor-Critic (HAC) reinforcement-learning framework for adaptive drone-assisted Internet of Things (IoT) communications. The framework jointly optimizes Drone Base Station (DBS) placement and user association under realistic wireless backhaul capacity and M/M/1 queueing-delay constraints. By separating slow-timescale DBS positioning from fast user association and coordinating both through a shared global reward, HAC enables scalable, lowlatency, and fairness-aware control in dense IoT networks. Extensive simulations over a $1000 \times 1000 \mathrm{~m}$ macrocell show that HAC reduces average end-to-end delay by roughly an order of magnitude compared to a greedy SINR-based policy and slightly outperforms a flat deep RL baseline, while improving loadbalancing fairness and maintaining queue stability under heavy traffic.
Understanding how social norms emerge and evolve on digital platforms is central to social computing. We present the Dynamic Norm Evolution (DNE) model, a cognitively grounded dynamical framework that integrates social reinforcement, confirmation bias, prestige influence, novelty-seeking, and memory decay into an evolutionary process of norm adoption. The model is formulated as a replicator-mutator system with cognitively modulated payoffs, yielding nonlinear differential equations that capture frequency-dependent selection among competing norms. Through fixed point analysis and bifurcation theory, we show that even simple two-norm systems exhibit bistable monopolization, tipping thresholds, and bias-driven transitions between dominant norms. Stochastic extensions reveal bimodal long-run adoption distributions and noise-induced switching between locked-in states. The DNE model provides a principled foundation for interpreting norm convergence, polarization, and sudden behavioral shifts in online ecosystems, offering a versatile theoretical lens for studying opinion dynamics, coordination, and collective behavior in socially networked environments.
