In this work, we propose a parallel low-depth flagged syndrome extraction protocol for fault-tolerant quantum error correction. The protocol combines parallel measurement of different stabilizers with dynamic selection of a minimal stabilizer set, guided by flag outcomes. Circuit-level optimizations, including CNOT reordering, are integrated while rigorously preserving fault-tolerance guarantees. Implementations for several distance-3 and distance-5 codes achieve substantial reductions in circuit depth and idle locations. Numerical simulations under a circuit-level depolarizing noise model demonstrate consistent pseudo-threshold improvements, such as 12.56% and 4.43% over prior schemes for [[5, 1, 3]], 16.2% for [[7, 1, 3]], and 14.55% to 30.13% for distance-5 codes, confirming the effectiveness of the co-optimized approach. This co-optimized framework offers a practical pathway towards more resource-efficient fault-tolerant quantum computation.
Power Distribution Networks (PDNs) face increasingly stealthy and causally complex cyber threats that challenge traditional detection approaches. While Large Language Models (LLMs) offer promising capabilities to reason about such threats, directly feeding fragmented, noisy, and heterogeneous machine logs to them often overwhelms their reasoning capabilities, obscuring critical causal chains. This paper presents MCCD, a Multi-Layer Collaborative Continuous Defense framework driven by an LLM-interactive Trusted Intelligence Agent (TIA). The TIA operates inside a TEE to ensure the authenticity and provenance of endpoint intelligence. A lightweight Temporal-Causal-Hierarchical Encoding (TCHE) transforms heterogeneous logs into structured Semantic Security Intelligence (SSI), enabling the cloud-based LLM to perform coherent causal reasoning. Through an adaptive closed-loop mechanism, the LLM iteratively refines the collected evidence and converges to a reliable final threat determination, rather than relying on one-shot inference. Evaluations on PDN-style datasets and DARPA Transparent Computing traces demonstrate that MCCD provides substantial gains in industrial power-system scenarios. These results demonstrate that combining Trusted Execution Environment (TEE), semantic intelligence, and iterative reasoning provides an effective and practical path toward robust threat determination in modern PDNs.
Dialogue summarization in informal conversational settings presents unique challenges, including fragmented syntax, colloquial expressions and dynamically shifting speaker roles. Existing summarisation models, largely adapted from monologue-oriented text, often fail to preserve these informal cues and neglect the discourse structures that underpin conversational coherence. To address these gaps, this study proposes DISCO-DIAL, a discourse and informality-aware neural framework for dialogue summarization. The framework introduces three novel components: (i) INFO-PRESERVE, an informality preserving preprocessing layer that retains colloquial elements, contractions and emotive signals such as emoji's, while adaptively filtering noise, (ii) Modified Bayesian Relevance Vector Machine (MBRVM), a discourse-aware graph neural encoder that models utterances as graph nodes connected through adjacency, speaker role and topic shift edges, thereby capturing both semantic salience and interactional dependencies and (iii) REL-AttnRank, a relevance guided attention ranking network that integrates hierarchical dual attention with Contrastive Relevance Loss and Semantic Fidelity Loss to ensure precise, semantically aligned sentence selection. The framework operates as an extractive summarizer, selecting salient utterances while preserving informal cues and discourse structure. Extensive experiments on the SAMsum benchmark dataset demonstrate that DISCO-Dial achieves ROUGE-1, ROUGE-2 and ROUGE-L scores of 87.5, 70.67 and 87.50, respectively, alongside a BERTScore of 99, significantly outperforming state-of-the art baselines. Qualitative analysis further reveals that the framework excels in maintaining dialogue intent, speaker relationships and contextual coherence, even in the presence of noisy, informal constructs. These contributions establish DISCO-Dial as a robust and scalable solution for real-world dialogue summarization with promising applications in conversational AI, customer support systems and meeting transcription.
Most existing studies on modulation recognition for orthogonal frequency-division multiplexing (OFDM) signals operate under the idealized assumption of perfect synchronization, thereby overlooking the practical challenges posed by real-world wireless environments. Such assumptions neglect the detrimental effects of timing and carrier frequency offsets, which inevitably arise in practice and can severely degrade recognition accuracy. In this work, we address this gap by proposing a novel modulation recognition framework tailored for coded OFDM signals operating under imperfect synchronization conditions over unknown channel coefficients. The core of the proposed approach is a maximum-likelihood (ML) formulation that jointly estimates the transmitted modulation format, the underlying channel response, and the synchronization parameters-namely, the timing and carrier frequency offsets. To efficiently solve this inherently coupled estimation problem, we adopt an expectation-maximization (EM) algorithm, which alternates between soft-data detection and parameter refinement. Unlike conventional schemes that rely exclusively on known pilot symbols for parameter estimation, our method leverages the soft outputs of a bit-interleaved coded modulation with iterative decoding (BICM-ID) receiver as virtual pilots. These decoder-generated reliability values are recursively fed back into the EM process, enabling continuous improvement of modulation recognition, synchronization, and channel estimates without requiring additional pilot overhead. Extensive simulation results demonstrate that the proposed scheme achieves superior modulation recognition accuracy compared to traditional synchronization-agnostic methods, particularly in scenarios with moderate-to-severe timing and frequency misalignments.
Automated generation of high-fidelity fonts for structurally complex scripts such as Korean and Chinese characters face significant challenges due to combinatorial glyph variations and nuanced stylistic details. While generative models like GANs and VAEs have advanced digital typography, they suffer from training instability, stroke distortions, and limited stylistic diversity. We introduce DML-Font, a diffusion-based framework that generates visually precise multilingual fonts (Korean, Chinese, English, numerals, including symbols) with unprecedented personalization. Our method employs phonetic-aware text encoding to decompose characters into linguistic components (Hangul jamo, Chinese strokes), ensuring structural integrity. Multi-reference style conditioning captures aesthetic nuances using diverse exemplars, while an iterative ResNet-based correction mechanism refines stroke alignment. Crucially, DML-Font leverages both standardized fonts and user handwriting samples to replicate personal writing styles. Quantitative evaluations demonstrate superiority printed fonts achieve SSIM 0.926 (Korean) and 0.920 (Chinese); handwritten fonts reach SSIM 0.861 (Korean) and 0.814 (Chinese), exceeding benchmarks by 6-15% across scripts. Concurrently, LPIPS (J,0.033 printed, J,0.081 handwritten) and FID (J,12.3 printed, J,21.2 Korean handwritten) confirm enhanced perceptual quality. The model generates complete TrueType Font sets with stylistic coherence across alphabets, numerals, and symbols. It also faithfully mimics user handwriting from minimal samples, bridging AI-driven design with cultural preservation. This work establishes a new paradigm for multilingual typography with applications in digital publishing, accessible communication, and heritage archiving.
Accurate detection and categorisation of coconut tree diseases are vital for enhancing agricultural output and sustaining farming practices. This study presents an advanced framework combining Enhanced YOLOv9 with EfficientNet for robust disease feature extraction. Our approach employs pseudo-labelling via Region Proposal Network (RPN) during the training phase to generate approximate bounding boxes for unannotated images. The EfficientNet backbone improves inter-class distinction and intra-class variability handling. During inference, the anchor-free detection head directly predicts disease region centers and dimensions without anchor priors, while Feature Pyramid Network (FPN) enables multi-scale feature aggregation for comprehensive detection of lesions ranging from small spots to large infected areas. Advanced data augmentation and a hybrid loss function enhance model generalization and detection capabilities. Experimental results demonstrate that our model achieves 97.34% accuracy, 99.8% precision, 94.67% recall, and a 97.17% F1 score, outperforming state-of-the-art architectures. This work presents an effective tool for field-based disease detection and classification, supporting timely interventions in coconut cultivation.
With the increase in demand for compact antennas appropriate for portable electronic devices has increased the implementation of fractal geometries in the field of antenna engineering. This survey provides a comprehensive review of fractal antennas, their practical applications, and the benefits they offer over traditional antenna designs, especially in terms of frequency. By combining self-similar structure, space filling attributes and multiband capabilities in fractal geometries not only make it compact but also support multi resonant antenna structure. Special focus is placed on the multi-resonance characteristics of these antennas, which are important for supporting the diverse frequency bands utilized in wireless communication. The survey begins with introduction to basic fractal geometries to generate different fractal structure. The Sierpinski and Minkowski fractal antennas support strong multiband behavior with bandwidth 3.1-10.6 GHz making them well suited for ultra-wideband and Wireless LAN applications. Koch and Hilbert curve antenna achieve 30-40% size reduction without compromising bandwidth making it suitable for portable and IoT devices with reduced gain 1.5-3.2 dBi. This review provides a comparative analysis of various antenna configurations, highlighting their merits and inherent limitations while identifying future research advancements with innovative design approaches to fulfill the demand of developing wireless communication applications.
This paper investigates the application of quantum principles-most notably superposition and entanglement-to reframe classical concurrency problems in software engineering. While other research has explored quantum concepts primarily as tools for parallelization, this work advances the discourse by proposing quantum-inspired abstractions for reasoning about mutual exclusion, race conditions, deadlocks, and the producer-consumer problem. By treating threads and shared resources as qubits and modeling their interactions via correlated quantum states within an abstract state-space, we present a framework that illustrates how quantum formalisms can conceptually model concurrency patterns, supported by mathematical formulations in Dirac notation and circuit representations. Superposition naturally models multiple execution paths, while entanglement provides symbolic global correlations rather than operational synchronization, offering new perspectives on longstanding concurrency challenges. Our simulations across four scenarios demonstrate that these quantum-inspired formulations can highlight order-sensitivity and contention patterns in a unified mathematical setting. These results do not constitute practical concurrency mechanisms, but they suggest how quantum principles may inform future hybrid quantum-classical analysis tools. We also conclude with a discussion on potential integration into hybrid systems, highlighting both practical benefits and open challenges.
The application of the Internet of Things (IoT) has been implemented for Agriculture 4.0 via the use of wireless sensor networks (WSN), including smart monitoring, supply chain management, agrochemical applications, smart water, and smart harvesting. An attacker trying to get into the Agriculture 4.0 network may employ a diverse approach to disrupt the operation of equipment that is grounded on the IoT, such as scanning attacks, distributed denial of service (DDoS) attacks, along with false data injection attacks. To identify unusual or suspicious behaviour and implement preventative steps regarding the intrusion risk, the IDS system keeps an eye on the network traffic. To better protect against cyber-physical systems, the internet of cars, smart grids, cloud computing, industrial SCADA systems, big data, smart grids, and wireless networks, some proposed intrusion detection systems have recently used deep learning (DL) approaches. Farmland Fertility Optimisation (FFO) with DL Enabled IDS for Smart Agriculture (FFODL-IDSSA) is a novel approach that the research proposes. This approach uses FFO for feature selection, GEO for hyperparameter tuning, and a Hybrid Deep Learning (HDL) model, which integrates BiLSTM with multi-head attention for accurate intrusion detection. Recognising intrusions in agricultural data accurately is the foremost focus of the proposed FFODL-IDSSA approach. The proposed approach uses numerous subprocesses, comprising feature selection, classification, and preprocessing, to accomplish this. To pick optimal feature subsets, the described FFODL-IDSSA approach uses an FFObased feature selection strategy. The detection and classification of cyberattacks are then carried out using the hybrid DL algorithm. In the DL models the tuning of the parameters is completed using the Golden Eagle Optimiser (GEO) method. To determine the enhanced performance of the offered technology, a comprehensive simulation is run. The proposed framework offers superior performance over traditional models, addressing the unique challenges of IoT security in agricultural environments.
Grover's algorithm utilizes quantum computing to find solutions in an unstructured search space with significantly fewer oracle calls-a square-root reduction compared to classical methods. Since each application of Grover's algorithm typically yields a single solution, discovering all M solutions in a space of size N requires at least M applications, resulting in O(VNM) oracle calls. To reduce this computational cost, we build on the observation that solutions in many problems are not uniformly distributed but tend to be densely populated in specific regions. Leveraging this observation, the proposed method adopts a quantum-classical hybrid approach: it uses classical computation to dynamically track whether such high-density regions exist during the application of Grover's algorithm, which is executed on a quantum computer. Upon identifying a promising subregion, the method restricts the subsequent search to that subregion. This targeted search enables more efficient discovery with fewer oracle calls. We analytically show that when the average solution density is rho in the subregions, the number of oracle calls can be reduced to O(M/ & check;p). Experimental results confirm that the efficiency gains become more substantial as N increases. For example, when N = 218, the baseline and proposed methods required 19K vs. 9.7K oracle calls, respectively; and for N = 224, the difference widened to 390K vs. 70K. We believe that this work lays the foundation for future research on discovering all solutions using Grover's algorithm more efficiently.
A high-precision semantic map that integrates environmental spatial information and terrain category semantics constitutes a key technology for the navigation of field robots in unstructured environments. Accurate real-time semantic segmentation in unstructured outdoor environments is still a fundamentally challenging issue, despite all the recent advancements in this field. To address this problem for field robots, this research proposed DMRINet, a semantic segmentation network appropriate to embedded platforms. With a backbone branch for global feature extraction, an interactive branch for information supplementation, and a residual branch for local detail compensation, this network employs an "encoding-decoding-parallel interaction" architecture. It can effectively balance segmentation accuracy and inference speed while adapting to complicated terrain scenarios. Finally, ablation and comparison analysis was performed using the HDU-Terrain dataset, a robot-perspective dataset with 4,000 frames of pixel-level annotations across several unstructured terrain types. Results show that the proposed DMRINet balances accuracy and speed, achieving 87.4% mIoU, 91.84% mPA, and 48.79 FPS inference speed, outperforming comparative models and demonstrating potential for application in diverse field robots.
To achieve quantitatively controllable desensitization intensity while preserving the statistical value of data, this study investigates a sensitive information desensitization method for juvenile delinquency data based on differential privacy. First, an improved gray clustering algorithm combined with a linear correlation coefficient is used to extract sensitive information from juvenile delinquency texts. Then, the differential privacy mechanism is applied by introducing Laplace distribution noise to desensitize the extracted sensitive data. To realize quantitatively controllable desensitization intensity, confidence analysis is employed to determine the scale parameter of differential privacy. Furthermore, the probability density function and distribution function of the Laplace noise are analyzed to regulate the privacy budget allocation, ensuring a balance between privacy protection and data utility. Through reasonable allocation of the privacy budget, the method effectively conceals sensitive information while maintaining the statistical characteristics of the original data, thereby improving the overall desensitization performance. Experimental results demonstrate that the proposed method can effectively extract sensitive information related to juvenile delinquency, achieving superior extraction performance. In addition, the desensitization process shows strong defense capability against sensitive information leakage, with perceptual loss remaining relatively high under different perturbation intensities (minimum value about 0.8). Meanwhile, the ambiguity degree under different data sizes remains low (maximum value about 0.1), indicating that the proposed method provides strong privacy protection while maintaining data usability.
Accurate classification of brain tumors using magnetic resonance imaging remains challenging due to the complexity of tumor morphology. While deep neural networks have advanced automated tumor classification, their large model sizes often limit their practical application. To address this limitation, we propose MobileNetV2 Attention Quantum Network (MAQNet), a hybrid quantum classical model based on the inverted residual blocks of MobileNetV2. MAQNet incorporates attention mechanisms and a variational quantum circuit to enhance feature learning. The model is designed to capture local and global contextual information while maintaining a low parameter count of only 1.3 million-an order of magnitude lower than that of most conventional convolutional neural networks (CNNs). We evaluated MAQNet against several established CNN baselines, including ResNet-50, MobileNetV2, EfficientNet-B0, DenseNet121, and InceptionNetV3, under identical training conditions for a fair comparison. The results revealed that MAQNet achieved an average accuracy of 97.98% over three training runs, exhibiting an excellent balance between predictive performance and parameter efficiency.
Effective resource allocation in cloud computing systems continues to be a significant problem, especially under constraints related to diverse resource types and stringent execution deadlines. This paper introduces an adaptive scheduling technique that improves system efficiency and task acceptance by jointly considering deadlines and resource types. It uses dynamic time-slot allocation based on arrival time, deadline proximity, and available processing units. When continuous allocation is not feasible, segmented execution ensures deadline compliance. Comprehensive simulation tests were performed with four workload sets consisting of 13, 22, 33, and 43 tasks. The proposed solution outperforms Haizea, Swapping, and CloudSim, achieving 89.5% task acceptance, over 95% utilization, and reduced rejections and rescheduling. This work's primary addition is its dual-focus scheduling approach, which combines deadline alignment with resource-type awareness, a combination seldom explored in existing literature. The algorithm produces excellent results with minimal operational complexity, making it a practical, scalable solution for Infrastructure-as-a-Service deployments. The current implementation of the proposed scheduling mechanism focuses exclusively on Central Processing Unit (CPU) resources. Other critical resources, such as memory, storage, and network bandwidth, are not yet integrated into the allocation process. Future work will extend the model to cover multi-resource allocation scenarios.
Static vulnerability analysis of source code is a critical research topic in cyber security. While manual and traditional automated methods exist, they struggle with efficiency and complex vulnerability types. Therefore, this study proposes CVDF (C/C++ Vulnerability Detection Framework), a novel static vulnerability detection framework based on Bi-LSTM neural networks designed to identify 6 types of vulnerabilities simultaneously. We construct a Vulnerability Eigenvector (VE) by extracting keywords and key operations from sliced, tokenized code. Furthermore, we redesigned the input, forgetting, and output gates of the Bi-LSTM network to specifically adapt to C/C++ static analysis. Experimental evaluations across multiple databases demonstrate that CVDF achieves an average accuracy of 98.4% on Buffer overflow vulnerability and an average precision of 91.7%, significantly outperforming traditional tools like Checkmarx while maintaining a low false-positive rate (3.2% for buffer overflows). Finally, we discuss the future development trends of CVDF, highlighting its potential to enhance modern software security practices. Our Code is available at: https://github.com/zhuln020/CVDF-A-Static-Vulnerability-Detection-Framework-for-C-C-Based-on-Bi-LSTM-Neural-Network.git
Recently, the amount of mobile data traffic transmitted by numerous Internet of Things (IoT) devices and user equipments (UEs) deployed in the coverage of cellular networks is rapidly increasing but cellular base stations alone cannot handle the traffic and provide uninterrupted service to users. To address this issue, 6th generation (6G) technology introduced unmanned aerial vehicles (UAVs) as aerial relay stations (ARSs) to provide improved service to UEs experiencing poor service performance from fixed ground base stations (GBSs). However, unlike GBSs that transmit signals directly, UAV-based relaying requires an additional wireless link for relay transmission, and the UE must be optimally associated with both the GBS and the UAV. In this paper, we propose a novel user equipment association (UEA) method, named expected-capacity-driven UEA (ECD-UEA), to maximize downlink system performance in multi-UAV-assisted cellular networks. The proposed method operates in two stages: first, classifying UEs based on received signal strength (RSS), and second, strategically reassigning UEs to UAVs by considering real-time load conditions and expected relay/access capacity. Simulation results demonstrate that the ECD-UEA method significantly outperforms baseline schemes, including Max-RSS and Log-Utility-Maximization (LUM) methods. Specifically, it increases the total system capacity by up to 11% compared to GBS-only networks and improves the capacity of the bottom 10% UEs by approximately 40% to 80%. Furthermore, the ECD-UEA method exhibits superior resource efficiency, achieving higher total throughput with fewer UAVs compared to conventional approaches. This confirms that the proposed strategy is a practical and cost-effective solution for balancing network loads and enhancing the quality of service (QoS) in dense 6G urban environments.
Alzheimer's disease (AD) is a chronic degenerative brain condition that leads to a gradual decline in memory, daily functioning, and thinking skills, mainly affecting the aging population. Accurate early-stage identification of AD is crucial for initiating appropriate treatments and improving patient management. Conventional diagnostic approaches depend on extensive clinical assessments and manual examination of neuroimaging scans, which are time-consuming, subjective, and prone to inconsistency. To overcome these limitations, recent advances in deep learning (DL) have emerged as an effective solution for automating AD and stage classification. This research proposes a novel modified TransNeXt, a hybrid DL architecture that integrates the local pattern recognition capabilities of convolutional neural networks (CNN) with broader contextual modelling offered by vision transformers (ViT). The proposed model analyzes brain MRI images to accurately categorize them into four stages of AD, such as moderate dementia, non-dementia, mild dementia, and very mild dementia. The model is trained and evaluated on the OASIS dataset, which offers a diverse and well-structured set of MRI scans. The integration of multi-head attention, depthwise convolution, and the convolutional gated linear unit (ConvGLU) improves the model's capability to capture subtle structural changes in brain MRI images. The proposed modified TransNeXt model achieves outstanding classification performance, obtaining 99.25% accuracy, 99.27% recall, 99.24% precision, and 99.25% F1 score. These results demonstrate the robustness of the proposed model as a reliable computer-aided tool, empowering clinicians with early diagnosis, comprehensive risk management, and well-informed treatment planning for AD.
Recently, methods that leverage the overfitting property of deep neural networks (DNNs) to enhance video quality while saving storage have been proposed. These approaches pair low-resolution (LR) videos with overfitted super-resolution (SR) models so the decoder reconstructs high-resolution (HR) videos. However, storing the model parameters incurs substantial overhead, which limits practical applicability. To address this, we propose a content-aware video overfitting method, Training the Significant Parameters (TSP). TSP combines gradient-based analysis with an adaptive overfitting strategy to improve SR reconstruction quality while reducing parameter storage. Specifically, TSP dynamically identifies the network layers that most affect performance via gradient analysis and updates only their parameters, improving parameter efficiency and compatibility. We further introduce a spatiotemporal overfitting strategy based on Hard Sample Learning (HSL). By allocating more training to texture-rich regions of a video, HSL improves the peak signal-to-noise ratio (PSNR) of SR results by 0.41 dB. Experiments show that, compared with the H.265 codec, TSP reduces data size by over 50% while achieving the same PSNR, and it outperforms stateof-the-art video codecs at low bit rates. TSP also adapts well to long, complex videos, demonstrating strong potential for practical applications.
In time-varying satellite-terrestrial integrated networks (STINs), rapidly changing link conditions and stochastic task arrivals make it difficult to minimize long-term task execution delay while maintaining queue stability. To shorten transmission distance and reduce latency, we deploy mobile edge computing (MEC) servers on low Earth orbit (LEO) satellites. Each ground task can be processed locally, executed on the serving LEO satellite, or relayed to a terrestrial cloud center. We formulate long-term delay minimization as a mixed-integer nonlinear programming (MINLP) problem with long-term queue stability constraints. To handle the inter-slot coupling caused by queue dynamics, we adopt Lyapunov drift-plus-penalty optimization and convert the original multi-stage problem into per-slot deterministic subproblems. Based on this framework, we propose DRLODE, an online computation offloading strategy using a dual-policy deep reinforcement learning architecture. We further introduce a ternary order-preserving action quantization mechanism to efficiently generate discrete offloading decisions. Simulation results show that DRLODE converges within about 2000 time slots and achieves an average task execution delay of about 0.1 s at a task arrival rate of 3 Mbps while keeping queues stable. Under a heavier load with a task arrival rate of 25 Mbps, DRLODE reduces the delay to about 3.4s, compared with about 11.6 s for the DDPGTO baseline and about 7.0 s for the ablation variant without the ternary order-preserving quantizer. These results verify the effectiveness of DRLODE in dynamic STIN environments.
Text-prompt-based artificial intelligence image generation plays a crucial role in visualizing concepts during early design and prototyping stages; however, existing approaches often exhibit limited semantic consistency between input prompts and generated images. To overcome this limitation, this study introduces a hybrid quantum computing-assisted deep learning framework for targeted text-to-image generation. The proposed method enhances prompt-image association by jointly modeling textual semantics and visual attributes using quantum-encoded representations. At the same time, deep learning is employed to validate and optimize the integration of objects, backgrounds, and attribute combinations. By improving semantic alignment at the representation level, the framework effectively reduces generation inconsistency. Experimental evaluation on publicly available text-to-image prompts and synthetic image datasets demonstrates that the proposed approach achieves a 11.4% improvement in similarity score and a 13.13% increase in F1-score compared with existing methods.
