With the rapid evolution of mobile connectivity, security threats such as distributed denial of service attacks have become a significant concern for network stability. This research proposes a defense framework that integrates zero trust architecture with a real-time monitoring system to enhance network protection. The system employs continuous identity verification and dynamic permission management to ensure that only authenticated devices can transmit data. By evaluating different machine learning approaches, the study demonstrates that a deep learning model focusing on sequence dependencies provides superior accuracy in identifying malicious traffic compared to traditional classification methods. Practical testing indicates that the proposed architecture effectively mitigates the impact of attacks and prevents the internal spread of threats, offering a more robust security solution than conventional network defenses.
With the rapid development of electronic technology, navigation and control technology, low-altitude flying targets have been increasingly widely used. Due to the low flight altitude, small size and high movement speed of low-altitude targets, these characteristics pose significant challenges to the tracking of low-altitude targets. How to achieve accurate and efficient tracking of low-altitude targets has become an urgent problem to be solved. To address this issue, this paper designs a low-altitude target detection and tracking method mainly from two aspects: the accuracy and real-time performance of low-altitude target tracking. Aiming at the requirements of real-time detection and tracking of low-altitude targets, two detection schemes based on YOLOv8 and YOLOv8-Ghost are discussed and analyzed. Through comparative experiments, the YOLOv8 algorithm exhibits excellent performance in low-altitude target detection due to its high accuracy, but it suffers from the problems of high power consumption and slow inference speed. Therefore, on the basis of YOLOv8, this paper integrates Ghost Convolution, which reduces power consumption and improves inference speed. Then, combined with the ByteTrack multi-object tracking algorithm, which consists of a target detection module, a trajectory prediction module and a hierarchical data association module, the detection and tracking of multiple low-altitude targets are realized. The datasets adopted in the experiments include low-altitude target images and video datasets (covering birds and unmanned aerial vehicles) [11]. The YOLOv8-Ghost algorithm is applied in the detection stage, achieving favorable detection results, with reduced model parameters, lower power consumption and significantly improved inference speed. The ByteTrack algorithm is used in the tracking stage, outputting visualized results with good tracking performance.
Mission-critical services, such as emergency communications, active safe driving (ASD), eV2X, and LEO satellite networking, require low-latency and reliable transmission in 5G-A TN) and 6G NTN LEO systems. In mixed-numerology operation, static BWP configuration and conventional scheduling may lead to vRB MX, RB fragmentation, and inefficient resource usage. To address these issues, this paper adopts an extended Sigmoid-based Cost-based Flow Scheduling (eSig) framework for queue-level flow scheduling. eSig evaluates each HoL packet by using a net-profit metric that combines QoS demand, vRB state, carrying cost, and service reward. The carrying cost is modeled by a dynamic exponential function, while the service reward is modeled by an adaptive extended Sigmoid function with MCS-aware and resource-aware factors. Simulation results show that eSig enhances vRB utilization and total net-profit while keeping delay within the required bound.
Telemetry generated by honeypots has a wealth of attack information but lacks in actionable semantics to inform triage decisions. While security experts usually get to deal with shell commands, protocol fields, partial payloads, and unsuccessful authentication attempts, triage calls for more abstract decision-making regarding the crucial attack mechanism used. This semantic mismatch poses a challenge for CAPEC inference and slows down decision-making. To solve the issue, we propose a decoupled pipeline that builds a behavioral profile out of raw data, uses an LLM in the organization’s premises only to determine the attacker’s intent, and finally retrieves potential CAPEC entries by running two simultaneous searches, semantic and lexical, and combining results with Reciprocal Rank Fusion. Evaluating our approach on 25 behavioral profiles, generated from data collected by Cowrie, Honeytrap, Dionaea, SentryPeer, and CiscoASA honeypots, yields the following metrics: P@1 = 0.760, MRR = 0.973, nDCG@5 = 0.741, Precision@5 = 0.632, Recall@5 = 0.511. An analysis of the ablative variants suggests that neither branch alone is capable of inferring attack intents accurately. In turn, compared to a monolithic frontier LLM, our system shows an acceptable performance trade-off and can be considered auditable, compact, and feasible for on-premises SOC deployment.
In conventional three-dimensional (3D) content production, multiple processes, such as modeling, texture settings, material adjustment, spatial arrangement, and rendering, must be performed separately using different tools. These operations require specialized knowledge of 3D computer graphics and virtual reality development, making it extremely difficult for nonexperts to create and experience 3D content. In addition, file conversion between processes, scale matching, and lighting settings must be performed manually, making production a complicated and time-consuming process. In this study, we developed a technical platform to automatically generate a 3D model from a single still image and visualize and experience the generated model in a virtual space. The proposed system was evaluated experimentally using the system usability scale with 50 paricipants, mainly students. The results were generally favorable.
In this study, we combined motion capture and virtual reality technologies to develop a dance education support system that allows players to efficiently learn by imitating the movements of a coach. The proposed system tracks the position of small sensors attached to various parts of the player’s body in real time. By comparing and analyzing the coach’s and player’s movement data in the same space, the system provides a mechanism to visualize and quantify the degree of agreement between the movements. This approach enables the identification of movement errors and differences in posture, which are difficult to detect with current learning methods using mirrors or videos. Consequently, this system allows players to objectively assess and refine their movements.
Wireless Mesh Networks represent a flexible and low-cost solution for providing reliable connectivity through self-healing nodes. However, the placement of mesh routers is a challenge and is classified as an NP-hard. In this paper, we utilize WMN-PSOHCDGA hybrid system to optimize node placement. We investigate the performance of Simplex Crossover (SPX) and Parallelotope-Shaped Blend Crossover (psBLX) methods in combination with three router replacement methods: Constricted Method (CM), Linear Decreasing V_max Method (LDVM), and Rational Decreasing V_max Method (RDVM). Simulations are conducted for medium-scale networks considering Boulevard client distribution where clients are arranged along a linear path such as a street or commercial avenue. Our results show that all configurations achieved full router connectivity. RDVM combined with SPX (SPX-RDVM) maximized client coverage, while CM combined with SPX proved better load balancing. We found that psBLX-RDVM provided better trade-off between client coverage and load balancing.
Multimodal federated learning (MFL) enables privacy-preserving collaborative training across distributed clients with heterogeneous data modalities by leveraging cross-modal representation alignment and fusion to integrate modality-specific information from the same physical process. However, these mechanisms are also prone to emerging security vulnerabilities, which can be exploited to implant stealthy backdoors into the global model. Existing attack methods either focus on unimodal settings or rely on input-level triggers, overlooking coordinated cross-modal backdoors in MFL environments. To address this gap, we introduce an attack model named CMBA that implants a coordinated cross-modal multi-level backdoor attack by manipulating multiple learning components, including the classification objective, representation alignment, and latent structure. This design allows the backdoor to remain semantically consistent across modalities while manipulating the underlying multimodal representation space, enabling the attack to persist under federated aggregation.
Archaeological cataloguing requires heterogeneous domain expertise and remains largely manual. This study presents a multimodal AI framework that decouples visual perception from typological classification through a two-stage architecture: a Visual Language Model (VLM) produces detailed visual descriptions of artifacts, while a dual Retrieval-Augmented Generation (RAG) system retrieves references from domain-specific knowledge bases, enabling grounded feature-to-type identification. Central to the design is the role of the domain expert, who contributes at four strategic junctures—knowledge base curation, contextual prompting, bounding box annotation, and output validation—ensuring that human expertise guides the process where it has the highest marginal impact. The system produces structured records designed for cataloguing interoperability and computational reuse. We validate the framework across three artefact domains (ceramics, architecture, and paintings) using sample images from the Pompeii excavation site, demonstrating the flexibility of our approach and the potential of multimodal AI for structured archaeological analysis.
It is obvious that stocking enough electricity for emergencies before disaster. In the case of Battery Swapping Electric Vehicle (BSEV) infrastructure, Battery Swapping Station (BSS) works as distributed energy resource (DER) at a disaster by utilizing its multiple charged spare batteries. However, it is difficult to improve energy stock performance under conventional operation on BSEV infrastructure where a BSEV swaps its battery if and only if its battery SOC (State Of Charge) falls below battery swapping threshold. This operation does not contribute to increase charging duration of spare batteries at a BSS. In this study, we discuss how to bring out BSS’s potential on charging ability by optimal battery energy storage management in cooperation with EVs for increasing energy stocks. For the purpose, this paper proposes a novel method named E-BCBS based on previous proposed BSS-centric battery swap approach, BCBS. The excessive BSS-centric approach on the proposed method, however, induces under utilization of BSEVs’ travel distance. The results of performance evaluation confirmed that E-BCBS well mitigates the under utilization of EV travel distance than the normal BCBS.
The explosive growth of micromobility has introduced a critical safety asymmetry in dense urban environments: while autonomous vehicles evolve into high-performance computing nodes, micromobility users constitute SWaP constrained (Size, Weight, and Power) endpoints, lacking the energy budget for advanced onboard sensing. This paper addresses this gap by proposing a Joint Radio Planning (JRP) framework tailored for Future 4D Mobility Networks. Unlike traditional vehicle-to-everything (V2X) approaches that rely on device-centric perception, we leverage 6G Integrated Sensing and Communication (ISAC) to transform the network infrastructure into a ubiquitous sensor. We formulate the planning problem as a maximization of the Network Utility Function ( U_net ), balancing the fundamental trade-off between communication rate and sensing Cramér-Rao Bound (CRB). Furthermore, we introduce a Rate-Distortion-Cost optimization model to mathematically demonstrate how network-centric sensing relieves the SWaP burden on vulnerable users. Finally, we detail a testbed-driven architecture designed for Urban Living Labs (ULL), incorporating Reconfigurable Intelligent Surfaces (RIS) as virtual anchors to eliminate blind spots in urban canyons.
The QR code, a two-dimensional symbol developed in 1994 and standardized by the ISO/IEC, has been widely adopted across consumer and industrial domains. Despite more than three decades of deployment, certain aspects of its specification remain insufficient to meet evolving application requirements. In consumer contexts, QR codes are used for web access via smartphones, yet native support for phishing-resistant mechanisms remains limited. In industrial settings, where QR codes function as data carriers, typical implementations offer inadequate provisions for data authentication and secure transmission of confidential information. To address these limitations, this study proposes a two-area two-dimensional symbol, QR Twin, whose design employs a chimera structure that overlays two QR codes within a unified symbol. The secondary area provides additional encoding capacity, enabling the integration of digital signatures, confidential data, and enhanced error-correction information. This study examines the structural configuration, functional features, and potential application domains of the QR Twin framework.
Having a layered abstraction of the internal structure of blockchain systems is valuable for system design, security analysis and comparison between platforms. Existing models in the literature range from 3 to 7 layers, with no agreement on naming or boundaries, and they do not consistently isolate essential architectural dimensions such as cryptography and execution. In this paper, we propose BL8, an eight-layer reference architecture for blockchain systems. BL8 introduces two innovations: (i) a dedicated cryptographic layer that formalizes cryptographic primitives as an autonomous and cross-cutting foundation, and (ii) a distinct execution layer that separates the execution environment from both consensus logic and application logic. We validate BL8 by instantiating it on Bitcoin, Ethereum, Hyperledger Fabric, Solana and Polkadot.
The increasing complexity and opacity of the online advertising supply chain have raised serious concerns about transparency and trustworthiness. In particular, dark pooling—a practice in which low-quality or malicious publishers covertly sell ad inventory by leveraging the seller identities of reputable domains—poses significant risks to both advertisers and brand safety. While prior work has demonstrated the prevalence of dark pooling in global ad ecosystems, its extent and economic impact in country-specific markets remain unclear. In this paper, we present the first empirical study of dark pooling in the Japanese web ecosystem. We address two research questions: (RQ1) How prevalent is dark pooling among Japanese websites? and (RQ2) Does dark pooling increase ad inventory prices? To answer RQ1, we construct a novel dataset by analyzing ads.txt and sellers.json files from Tranco Top 100K websites with the .jp domain, and classify pooling structures based on organizational relationships inferred from Whois data. To answer RQ2, we collect real-world bidding data using Header Bidding (HB) and analyze the impact of pooling types on cost-per-mille (CPM).
Bi-objective fuzzy system optimization must balance predictive performance and interpretability. But most studies assess the search process only through the final Pareto front or a selected solution. This paper presents FuzzyTrace, a post-hoc attribution framework for standard Non-dominated Sorting Genetic Algorithms-II (NSGA-II) search on a fixed-rule Mamdani Type-1 fuzzy system. FuzzyTrace logs accept search events and export three process views: the Operator Attribution Matrix (OAM), the Population Evolution Graph (PEG), and the Convergence Driver Score (CDS). CDS is the main quantitative view and allocates the Hypervolume (HV) gain at each generation in proportion to operator activity. Experiments on two regression benchmarks with 30 independent runs have shown that the induced operator rankings remained stable across runs (mean consensus Spearman ρ≥ 0.87 ). The weighting ablation showed that moderate changes in the interpretability term mainly changed the geometry of the second objective, while the selected predictive performance and the consensus attribution hierarchy stayed similar.
Unlearning is a technology that removes the contribution of personal data to the learned model by reflecting a specific individual’s deletion request, but since de-identified data such as k-anonymity are often used for learning, the re-identification risk changes that unlearning affects individuals other than those subject to deletion have not been sufficiently identified. This study applied quasi-identifier-based k-anonymization to the adult census dataset and then analyzed the effect of unlearning on the individual re-identification risk distribution by learning model. The risk distribution for each individual's re-identification was calculated through a shadow model-based membership inference attack, and risk migration was quantified through the risk change and risk increase rate before and after unlearning in the training data after deletion. As a result of the experiment, it was confirmed that risk migration for each learning model is issued when unlearning is performed on the data to which de-identification is applied.
This paper addresses the issue of effectively building domain ontologies for consent management, since traditional manual methods frequently encounter knowledge acquisition impediments. To overcome these limitations, we leverage Large Language Models (LLMs) and Natural Language Processing (NLP) techniques. Our technique uses Gemini to automatically extract out important concepts and relationships from a collection of consent forms. The data are then imported into Protégé, where domain-specific rules are added to the ontology schema. Through OWL transformation and OntoGraph visualization, users are presented with a graphical representation of the ontology. Our approach is evaluated on 53 consent forms related to social media exposure and parental consent. The findings show that the approach reduces data needs and dependence on domain-specific expertise while having equal levels of quality and accuracy.
Homomorphic encryption (HE) enables privacy-preserving machine learning inference by allowing computation directly on encrypted data. However, due to the high computational overhead caused by encryption and ciphertext operations, there are still limitations in the practical use of homomorphic encryption-based intrusion detection systems (IDS). Therefore, this paper proposes a hybrid SIMD packing technique to optimize the inference performance of a logistic regression IDS model in the CKKS homomorphic encryption environment using a real-world dataset consisting of 48 network traffic features. The proposed technique reduces the number of ciphertexts by partitioning the 48 features into groups of 8, and minimizes the number of rotation operations by applying a log-based rotation-and-sum strategy. It maintains the same detection accuracy as plaintext while achieving about a 56× latency improvement over Single-wise HE and a 2.5× improvement over Batch-wise HE.
Food waste reduction and sustainable composting remain important environmental goals. Traditional composting relies heavily on manual monitoring, which often leads to inconsistent conditions and slow decomposition. This study presents an IoT-based Smart Compost robot designed to automate environmental sensing and basic regulation inside a compact compost chamber. The system uses air and soil sensors with one-minute updates based on the average of twenty readings, while the fan, mist sprayer, mixer, and LED module operate under a simple rule-based control strategy. Experimental results have shown that the system collected data reliably but was unable to maintain good composting conditions. Air and soil temperatures remained between 20 and 30 ^∘C , below 40–60 ^∘C required for thermophilic composting, and humidity levels fluctuated widely. These outcomes suggest low microbial activity and highlight the challenges of heat retention and moisture control in small indoor composting systems. Overall, the findings demonstrate the feasibility of automated monitoring while identifying key limitations. Improvements of insulation, airflow design, and control algorithms are needed to achieve more stable and efficient composting in compact indoor environments.
Early detection of pancreatic cancer using computed tomography (CT) is challenging due to its small size, variable shape, low contrast and invisible signs of tumours. This work introduces a two-stage deep learning pipeline: the first stage uses a U-Net model, which is a symmetric encoder–decoder network. The model uses skip connections for pixel-level pancreas segmentation, which produces binary masks that isolate the organ from the remaining CT backgrounds. The second stage uses a pre-trained ResNet-18 convolutional neural network on the segmented regions to classify the pancreas as cancerous or non-cancerous. The model produces pancreas masks, which are easy to understand, and generates cancer probability scores. The U-Net achieved a segmentation accuracy of 99.72
