Abstract The Internet of Electric Vehicles (IoEV) has emerged as a key component of future networks. However, some computation‐intensive vehicular applications cannot be executed locally owing to IoEV infrastructure limitations and computing resource bottlenecks. In this study, we propose a task offloading and resource management scheme based on Parked Electric Vehicle (PEV)‐assisted distributed edge intelligence (DEI), making use of underutilized PEV resources to handle offloaded tasks. In the proposed scheme, distributional reinforcement learning, the normalized average bargaining solution (NABS), and V2G charging scheduling are jointly combined to dynamically control availability prediction, resource sharing, and scheduling. The proposed scheme maximizes hybrid optimization benefits through PEV and edge server cooperation. Simulation results confirm performance improvements of 10%, 10%, and 15% in normalized service payoff, system throughput, and task failure rate, respectively, compared with existing benchmark protocols. Open issues and research directions for PEV‐DEI systems are also discussed.
Abstract In this study, we propose a priority‐based contention‐avoidance (PCA) resource‐allocation method for disaggregated data centers (dDCs) that achieves a round‐trip time (RTT) of under 1 μs. The proposed scheme employs the disaggregated resource manager (dRM), which applies both the contention‐avoidance method and the best‐fit allocation (BFA) method, depending on whether an application is memory‐intensive (MI) or CPU‐intensive (CI), to ensure lower memory‐access delays for delay‐sensitive, higher‐priority services while promoting efficient resource allocation. This approach reduces contention and maintains priority balance by minimizing traffic concentration and prioritizing higher‐priority services across specific memory locations. We evaluate the performance of the proposed scheme in terms of RTT, memory‐access queuing delay, and end‐to‐end (ETE) delay using the OPNET simulator. The results demonstrate that the PCA scheme achieves the lowest latency, with a maximum of 0.89 μs, compared with maximum latencies of approximately 2.6 and 1.2 μs for the first‐fit allocation (FFA) and BFA schemes, respectively, when the offered load is 0.9.
Abstract We enhance the performance of SMAUG‐T, the winning algorithm of South Korea's Post‐Quantum Cryptography competition, by applying an NTT‐friendly ring homomorphism tailored to x86/64 and AVX2/AVX‐512. This paper makes two main contributions: (i) We integrate the NTT into the reference C(x86/64) implementation of SMAUG‐T and apply a more efficient NTT‐friendly ring homomorphism optimized for parallel processing. Previous attempts have also been made to apply the NTT to SMAUG‐T on AVX2; however, they have not addressed the methodology for selecting optimal primes in close conjunction with implementation optimization. (ii) We redesign several state‐of‐the‐art PQC optimization techniques, including layer merging, lazy reduction, shuffle, and negative wrapping convolution, to suit SMAUG‐T in parallel environments. Our C‐based x86/64 implementation achieves a performance improvement on matrix‐vector multiplication compared with the reference, whereas the AVX‐512 assembly implementation yields a improvement over AVX2 in the same operation.
Large language models are vulnerable to prompt injection because transformer attention treats tokens too uniformly across system and user roles. We propose the adaptive orthogonal role-aware transformer (AORT), a lightweight retrofit that improves role-sensitive decoding. AORT preserves the base token embeddings and adds at each transformer layer a learned orthonormal re-encoding and a role-gated additive attention bias to maintain separation between instruction and user representations during context mixing. We evaluate AORT on direct and indirect prompt-injection benchmarks, including the BIPIA-text benchmark. AORT achieves the lowest attack success rate (ASR) in most settings and successfully resists all eight evaluated attacks on LLaMA 3.1-8B. AORT reduces the ASR from 13.6% to 2.1% on LLaMA 3.1-8B evaluated with the BIPIA-text benchmark. These findings indicate that explicit role-priority structure in the forward pass can substantially improve prompt-injection robustness with minimal architectural overhead.
This study presents an empirical investigation of propagation loss characteristics at 255 GHz in indoor office and industrial environments, focusing on the directional beamforming performance under varying beam alignment conditions. Using a wideband channel sounder equipped with Tx and Rx horn antennas, three beamforming schemes are analyzed: (1) the optimal best-beam pair, (2) -th ordered beam pairs, and (3) randomly selected beam pairs. While the best-beam propagation loss is well characterized by conventional distance-dependent models (alpha-beta and close-in propagation loss models), the -th ordered and random beam pairs exhibit converging excess-loss behaviors. Notably, line-of-sight (LoS) links incur substantial penalties under random beam alignments, whereas non-LoS links incur smaller penalties, reflecting the inherent multipath diversity available in indoor channels. These findings provide experimentally grounded insights into the trade-offs between beam management overhead and system performance, establishing a foundation for interference analysis and beam management design of future sub-THz wireless networks.
Abstract This study verifies the feasibility of surface‐wave communication as a solution to the severe attenuation and multipath fading encountered by free‐space wireless systems in metallic environments. Surface‐wave characteristics on dielectric‐coated metal surfaces are analyzed in terms of field distribution and attenuation with respect to dielectric thickness and frequency. Experiments conducted in a structure emulating a watertight door show that placing a dielectric layer on the metal surface increases the received signal strength by about 5 dB and improves throughput by more than fourfold compared with the case without the dielectric layer. Additional experiments comparing surface‐wave and free‐space communication in the same ship‐like structure demonstrate that surface‐wave communication achieves approximately a 7 dB increase in received signal strength and more than a 15‐fold improvement in throughput relative to free‐space propagation. These results confirm that surface‐wave techniques can significantly enhance wireless coverage and reliability inside dense metallic structures and provide practical insights for future industrial wireless communication system design.
We introduce a hybrid reinforcement learning (RL) framework for dynamic budget allocation that blends Dirichlet-inspired stochasticity with quantum-mutation genetic refinement. Trained on Apple Inc.'s quarterly financials (2009-2025), the RL agent learns to allocate budgets between R&D and SG&A to maximize profitability, whereas penalties enforce adherence to historical spending patterns and prevent unrealistic deviations. A Dirichlet state-evolution model that shifts financial contexts and a genetic algorithm augmented by quantum mutation implemented as parameterized qubit-rotation circuits refine the policy to escape local minima and improve generalization. Generation-wise rewards and penalties are logged to visualize convergence and policy behavior. On held-out fiscal data, the proposed approach achieves near-perfect alignment with actual allocations (cosine similarity ; KL divergence ), underscoring the promise of combining deep RL, stochastic modeling, and quantum-inspired heuristics for adaptive enterprise budgeting.
This paper presents an extended-parity-based rateless polar coding scheme, which is designed to enhance adaptive transmission performance in broadband free-space optical channels. The proposed rateless polar codes are extended to both blocklength and coding rate variations, enabling finer adaptability to time-varying channel conditions. The proposed design leverages repetition-based XOR operations to strengthen the polarization effect and maintain reliable decoding as incremental parity symbols are transmitted. Our scheme is specifically investigated for downlink communication from high-altitude platform stations to ground stations, where turbulence-induced fading, pointing loss, and orientation divergence significantly degrade optical link reliability. Simulation results demonstrate that our extended-parity rateless design achieves notable performance improvements and enhances robustness against turbulence and misalignment effects, confirming its potential for next-generation, non-terrestrial optical communication systems.
The use of advanced imaging technologies has enabled the generation of high-resolution images in the medical domain. However, for efficient processing with limited computational resources, such images are generally processed as groups of small sub-images or patches. In this study, we propose a generative neural network that, when trained using a sequence of histological images, is capable of reconstructing the sequence when required. We demonstrate that this model, trained on sub-images of a high-resolution image, can serve as an alternative representation of larger images because it reconstructs the original image as a sequence of sub-images. Experimental results on images collected from the Grand Challenge breast cancer dataset demonstrate excellent PSNR, SSIM, and UQI values for the high-resolution images reconstructed by the proposed model. When analyzing the latent space of the models for different cancer images, individual images appear to occupy specific points within the space, which are distinguishable but non-separable, indicating that the proposed model is capable of capturing and reconstructing the structural features of individual images.
Model parameter leakage transforms black-box settings into near-white-box threats, enabling the insertion of highly transferable adversarial examples through exposed model internals. Defending against such attacks requires fast, robust, and minimally invasive adaptation methods. To address this challenge, we introduce a lightweight plug-in defense module that refines corrupted latent features while preserving a frozen backbone model. Our approach leverages ordinary neural differential equations with a Hopfield-type architecture, which is designed to purify latent features with strong stability properties. The key innovation of our model is a channel gain normalization technique that analytically binds the Lipschitz constant, thereby improving training stability and robustness. Notably, our method requires only 5% of the original training data to outperform standard adversarial training on both clean and adversarial samples, achieving 84.8% robust accuracy on CIFAR-10 (+6.2% over adversarial fine-tuning), while maintaining a clean accuracy of 90.4%. Furthermore, our model effectively adapts to repeated model leakages and memory-aware attacks by supporting rapid reinitialization. Extensive experiments on the CIFAR-10 and CIFAR-100 datasets demonstrate the effectiveness of the proposed approach. This work represents a practical step toward fast, computationally adaptive defenses against model leakage attacks, which are becoming increasingly common in modern machine learning environments.
This paper proposes a hybrid satellite-terrestrial relay network that integrates advanced technologies, such as physical layer security, fountain coding, non-orthogonal multiple access, reconfigurable intelligent surfaces (RISs), and partial relay selection (PRS). The satellite transmits fountain-coded packets to terrestrial relays, where the best relay selected using the PRS algorithm employs the decode-and-forward technique to deliver data to two user clusters with RIS support. This setup enhances both reliability and security. Meanwhile, eavesdroppers attempt to intercept the RIS-reflected signals. Exact closed-form expressions are derived for the outage probability (OP) and system OP of legitimate users, as well as the intercept probability (IP) and system IP of eavesdroppers. Monte Carlo simulations confirmed the analytical results and examined the effects of the key parameters. Additionally, two benchmark models were developed to facilitate performance and security comparisons with the proposed scheme.
Brain tumor is an abnormal growth of cells within the brain and a life-threatening neurological condition worldwide. Rapid cell growth is associated with higher mortality rates. Hence, timely and accurate segmentation and classification are essential for improving survival. However, variations in brain anatomy complicate accurate tumor boundary localization and glioma classification. In view of this, the present study proposes density peak guided superpixel fusion (DPGSF) with stacked long short-term memory (SLSTM) using fluid-attenuated inversion recovery images from the BRATS datasets. The DPGSF integrates density peak clustering with simple linear iterative clustering and Felzenszwalb superpixels to accurately segment brain tumors. Modified anisotropic diffusion is applied in the preprocessing stage to enhance edge preservation and noise suppression. The SLSTM utilizes clipped rectified linear unit activation to improve feature dependency modeling. DPGSF-SLSTM achieves classification accuracies of 99.16%, 99.66%, and 99.86% for BRATS2018, BRATS2019, and BRATS2020 datasets, respectively, outperforming state-of-the-art methods including ResAttU-Net and Caps-VGGNet.
This study proposes a geometry- and color-aware 3D data-augmentation framework to enhance fashion asset segmentation for digital twin applications. This study focuses on converting 2D fashion videos into 3D point-cloud data and augmenting the minority classes. The constructed dataset comprised 502 mannequin-wearing scenes and additional single-asset captures, totaling 628 instances across 16 categories and 104 items, all recorded using an iPhone 14 Pro. To address class imbalance, three augmentation strategies are introduced: SMOTE3D (normal-aware coordinate interpolation with red-green-blue jitter), scene reconfiguration, and class-consistent color shifts. Using OctFormer, OA-CNNs, and PTv3, the augmented data consistently improved the accuracy, intersection over union (IoU), and mean average precision (mAP), yielding significant gains over a Mix3D plug-in baseline, as confirmed by a paired Wilcoxon test.
Sequential recommendation is widely used in domains such as e-commerce and social media but still suffers from data sparsity, long-tail distributions, and popularity bias. Large language models (LLMs) provide rich semantics and transferability, yet their integration with collaborative signals remains challenging due to representation inconsistency, high prompting costs, weakened user modeling, and unresolved bias. To address these issues, we propose LLMSQRec, a unified framework that fuses LLM-derived semantic features with ID embeddings via dual-view modeling. Specifically, representation inconsistency is alleviated through semantic-collaborative alignment; prompting costs are reduced by combining frozen LLM embeddings with lightweight vector quantization; user modeling is strengthened by jointly encoding semantic and collaborative signals, and popularity bias is mitigated through curriculum learning with popularity-aware regularization. Experiments on multiple real-world datasets demonstrate that LLMSQRec significantly outperforms existing methods in sparse and long-tail settings, achieving superior accuracy, robustness, and diversity.
Aspect-based sentiment analysis focuses on discovering the sentiment polarity of specific aspects within textual data. The subtle complexities of language pose a major challenge, and identifying multiword aspects and their associated sentiment polarities hinders the development of effective models. To address the limitations of supervised approaches, unsupervised clustering can be employed to group similar aspect features. We propose an unsupervised hybrid architecture called the BERT + GPT-2 fusion model. Our main technical contributions include aspect extraction utilizing the BERTopic model. Our approach is evaluated utilizing silhouette and coherence scores. To identify the optimal number of aspect clusters, cluster quality is measured by calculating the average silhouette and coherence scores. The clusters that maximize the average values are well-defined clusters. The identified aspects are then used to train the BERT+GPT-2 fusion model. The proposed technique was evaluated on the 515 K Hotel Reviews Data dataset, demonstrating its efficacy for sentiment analysis. Our fusion model outperformed state-of-the-art approaches, including GPT-2 and convolutional neural network baselines.
Accurate feedback of downlink channel state information (CSI) plays a vital role in enabling beamforming in frequency-division duplex, massive multiple-input multiple-output (MIMO) systems. However, the feedback overhead scales with the number of antennas, severely impacting uplink efficiency. While transformer-based methods, such as TransNet, have exhibited high-reconstruction accuracy, their computational complexity limits the practical deployment on user equipment. STNet reduces this complexity via spatially separable attention but still underperforms TransNet in terms of accuracy. In this letter, we propose SPINNet, a lightweight CSI feedback network that combines spectral and inception modules to extract efficiently frequency-domain features. Experimental results on the COST2100 dataset show that the proposed SPINNet outperforms STNet by only half its computational cost and surpasses TransNet in certain scenarios.
File fragment classification is an essential task of identifying the file type given an incomplete binary file fragment. Despite their importance, existing deep-learning methods face two key limitations: (1) extensive hyperparameter tuning for varying fragment sizes and (2) considerably degraded inference performance when tested on fragment sizes different from those used during training. To address these challenges, we propose the byte-segment convolution and attention network (ByteSCAN) that introduces a fixed-slicing approach. Fixed slicing divides the binary data into overlapping segments of uniform size, enabling consistent feature extraction with a moderately deep convolutional neural network (CNN) and efficient feature integration through a self-attention layer. Experiments on two public datasets show that ByteSCAN outperforms previously published methods in terms of competitive training time and efficient inference while also demonstrating adaptability to varying fragment sizes. The code is available at https://github.com/SpatialAILab/ByteSCAN.
The early and precise identification of coronary artery disease (CAD) is critical for effective treatment and improved patient outcomes. This study offers a novel approach for enhancing CAD detection by combining quantum computing with advanced neural network techniques. We introduce a quantum-enhanced recursive feature elimination (QE-RFE) method that utilizes quantum processing power to optimize feature selection more effectively compared with traditional methods. The QE-RFE technique is integrated with a physics-inspired neural network (PINN) that leverages domain-specific knowledge of cardiovascular physiology and improves diagnostic performance. The PINN framework is constructed based on the physical principles of cardiovascular dynamics and offers a stable and interpretable model for analyzing medical data, yielding enhanced diagnostic accuracy and precision. Experimental results demonstrate that the proposed model achieves an accuracy of 98.26%, significantly outperforming state-of-the-art techniques and demonstrating enhanced computational efficiency. The proposed system represents a significant advancement in CAD detection and has promising clinical applications.
As we move toward the era of 6G networks, the emergence of numerous end-to-end services with diverse user demands is anticipated. To support these services under varying network conditions, traffic must be routed through optimal paths that satisfy quality of service (QoS) requirements while minimizing transmission costs. Furthermore, climate change concerns are increasing pressure to reduce both energy consumption and carbon emissions resulting from service operations. Given that these complex factors affect multiple domains, it is essential to develop an effective method for routing optimization. To address this issue, we propose a metaheuristic optimization method for end-to-end service routing that considers dynamic network metrics and computing site information to reduce energy consumption and carbon emissions. Evaluation results show that our approach selects near-optimal paths by accounting for various factors, including QoS, energy consumption, and carbon emissions. Compared with benchmark schemes, our model reduces the joint energy and carbon objective by up to 63%, average service latency by up to 75%, and maintains the highest availability across all scenarios.
This paper proposes a predictive handover decision framework for optimizing handover planning in integrated low Earth orbit (LEO) satellite-terrestrial networks. To forecast the fluctuations in received signal reference power (RSRP) caused by the high mobility of mobile terminals (MTs) and satellites accurately, we employ a convolutional neural network-long short-term memory encoder-decoder architecture. This forecasting model enables each MT to predict its future RSRP levels independently over a given time horizon. By leveraging the predicted RSRP trajectories, each MT autonomously determines its handover strategy, including its optimal target and timing. To enhance the performance of these decentralized handover decisions, we integrate a multi-agent deep reinforcement learning approach based on the QMIX algorithm. This approach enables coordinated policy learning among MT agents while preserving decentralized execution, thereby improving overall system stability and handover efficiency in highly dynamic environments.
