Vehicular networks support high-frequency data exchange for autonomous driving, but they demand dynamic bandwidth allocation. This drives the development of adaptive bandwidth data-aware technology. However, traditional adaptive bandwidth data-aware techniques face issues like imprecise bandwidth allocation and insufficient data accuracy. This study proposes an adaptive bandwidth data-aware technology based on spectrum allocation to improve data sensing accuracy. The technology integrates an adaptive merge strategy into the fireworks algorithm by dynamically adjusting explosion range and merge probability to optimize spectrum allocation. Experimental results show that the technology achieves a sensing accuracy of 99.0%, with a maximum adaptive adjustment time of only 243 ms. In real-world operation, system stability and data security reach 90.0% and 98.3% respectively, significantly outperforming comparative techniques. These results demonstrate that the proposed technology delivers excellent accuracy and efficiency in bandwidth data sensing. It effectively addresses the current challenges of low accuracy and inefficiency, offering new ideas for bandwidth data sensing and supporting the intelligent and efficient development of autonomous driving.
Smishing attacks pose an asymmetric-risk problem, where missing a threat causes significantly greater damage than false alarms. While transformer-based models achieve robust classification performance, their computational cost hinders continuous inference on resource-constrained edge devices. This paper proposes a lightweight, cost-sensitive deep learning framework for detecting smishing and suspicious SMS messages on edge devices. We formulate Weighted Binary Cross-Entropy (WBCE) as empirical risk minimization under a reweighted probability measure, combining a Bidirectional Gated Recurrent Unit (BiGRU) encoder with Masked Global Pooling to strictly exclude padding noise from pooled representations. Experiments on the UCI and LSDST-2022 datasets confirm that TF-IDF + Logistic Regression remains a robust sparse baseline in-domain. Our BiGRU+MaskedPool+WBCE model achieves high recall and competitive Average Precision, ensuring the highest recall among neural baselines. Under a harmonized cross-domain setup, it delivers higher recall than TF-IDF + LR in both transfer directions. Deployment profiling shows that our model requires only a 5.21 MB serialized footprint and 0.25 ms CPU latency, compared to 256.02 MB and 40.78 ms for DistilBERT. Finally, cost analysis confirms that operational thresholds must be calibrated to match the deployment-specific error penalties.
Unmanned Aerial Vehicles (UAVs) have become vital components in modern wireless systems, particularly within Wireless Sensor Networks (WSNs) and Internet of Things (IoT) ecosystems, where timely and flexible data collection is critical. Efficient trajectory planning is essential to balance mission duration, energy consumption, communication reliability, and service continuity. While traditional models such as the Travelling Salesman Problem (TSP) have been widely used, they often fall short in modelling realistic UAV missions. In contrast, the Close Enough Travelling Salesman Problem (CETSP) provides a more practical framework by allowing UAVs to collect data within a defined communication radius rather than requiring direct contact with sensor nodes. In this study, we perform a comparative analysis of four advanced UAV path planning strategies: CETSP-M, CETSP, TSPN, and TSPN-M. Each method is evaluated in conjunction with metaheuristic optimization algorithms, including Genetic Algorithm (GA), Artificial Bee Colony (ABC), and Ant Colony Optimization (ACO), to assess their effectiveness in addressing complex routing challenges. The evaluation focuses on minimizing travel distance across various scenarios involving 25, 50, 100, and 150 target areas. The results indicate that CETSP-M consistently outperforms the other approaches in terms of path length reduction and convergence performance, especially under varying node densities and communication radii. Furthermore, this study standardizes key performance metrics and provides application-driven insights to guide the selection of suitable algorithms for use cases.
Drug recommendation systems (DRS) play a crucial role in personalized medicine. Traditional Collaborative Filtering (CF) methods often face challenges with data sparsity and cold-start problems due to their reliance on drug/user identifiers (IDs). To address these limitations, we propose HyBLRecs, a recommendation system leveraging content analysis and reliability enhancement. HyBLRecs eliminates the reliance on IDs, instead employing a three-branch hybrid network architecture to deeply mine content from patient reviews. The model integrates the semantic capabilities of BioBERT, sentiment analysis, and statistical features to train a high-performance LightGBM classifier (AUC approximate to 0.98). Finally, an efficient confidence-based ranking mechanism is applied to balance predictive efficacy, stability, and evidentiary reliability (N). Experimental results on the Drugs.com dataset demonstrate that HyBLRecs outperforms existing methods, achieving an F1 score of 96.91% and an overall accuracy of 95.63%, which confirms its strong generalizability even for unseen drugs.
Missing values are a pervasive issue in real-world datasets, often degrading predictive performance and introducing computational instability. Existing imputation approaches typically focus on either data distribution modelling or attribute correlation analysis independently, limiting their ability to capture the underlying data structure comprehensively. Furthermore, most current methods are supervised and require complete datasets for training, restricting their applicability to fully missing or large-scale datasets. To overcome these limitations, this paper proposes kTails-WGAN, a novel unsupervised imputation framework based on Wasserstein Generative Adversarial Networks (WGANs). The proposed method incorporates data probability distributions into the discriminator to improve imputation accuracy and introduces a refined loss function to better preserve distributional consistency. Unlike conventional supervised approaches, kTails-WGAN does not require complete datasets during training, making it particularly suitable for datasets with extensive missing values. Experimental evaluations on benchmark datasets demonstrate that kTails-WGAN achieves competitive and, in many cases, superior performance compared with existing imputation methods, as measured by RMSE and AUC-ROC metrics.
Heart failure (HF) is a major cardiovascular condition that requires early and accurate diagnosis to improve patient outcomes. Echocardiography is widely used for cardiac assessment, however, the analysis of echocardiography videos using deep learning models often suffers from high computational complexity and redundant temporal information, limiting applicability in real-time clinical environments. To address these challenges, this study proposes a FAN-enhanced Vision Transformer integrated with distributed bidirectional optical flow for automatic heart failure detection from echocardiography videos. The proposed method introduces a preprocessing pipeline consisting of a frame refocusing algorithm to suppress stationary pixels, distance-based frame selection to reduce redundant frames, and distributed bidirectional optical flow to pad frame sequences. Furthermore, the Vision Transformer architecture is enhanced with Fourier Analysis Network (FAN) layers, enabling the model to capture periodic cardiac motion patterns while maintaining computational efficiency. Experimental evaluations were conducted on three standard echocardiography views. The proposed method achieved its best performance on the A2C view with an accuracy of 95.95% and a normalized Matthews Correlation Coefficient (MCC) of 96.02%. This result is statistically significant and better than other state-of-the-art methods. This approach provides a more robust and efficient diagnostic, as the FAN layers capture cardiac periodicity while requiring fewer parameters.
Attribute encryption technology presents increasing challenges to the scalability of IoT systems as the number of devices and data volume grow. This study designs a scalable attribute encryption method for IoT semi-structured data using an enhanced recurrent neural network (RNN). The ant colony algorithm optimizes queries on semi-structured data, generating efficient query plans. A backpropagation (BP) neural network learns similarity feature vectors, eliminates redundant data through batch processing, and applies optimal classification matrices to reduce feature dimensions. Additionally, a recursive neural network with multiple hidden layers extracts hierarchical abstract features, capturing complex patterns from semi-structured data. For encryption, a quantum cellular neural network’s chaotic generator dynamically produces keys, while XOR operations and balance coefficients enable adaptive encryption and storage. Experimental results demonstrate that the proposed method achieves an encryption strength of approximately 0.98 (normalized to 1.0), operates within 150 ms, and maintains low memory consumption. Moreover, the ciphertext ASCII codes exhibit a uniform distribution, providing resistance against frequency-based plaintext and ciphertext attacks. By integrating machine learning, optimization algorithms, and quantum-inspired encryption, this framework addresses scalability and security requirements in large-scale IoT deployments, ensuring efficient data processing and robust protection against evolving threats.
Classification models have emerged as the primary tools for numerous automatic computer vision tasks. However, they are susceptible to adversarial attacks, that can be harmful, but can also be employed to protect private information from classification-powered threat models designed to extract data from images. Black-box attacks, where an attacker has no knowledge about the model, are the most challenging ones, but also the most realistic ones. The difficulty is furthermore increased when one intends to create high-resolution adversarial images. We introduce NbuGAN, a novel black-box attack that creates high-resolution adversarial images deceiving image classification models in the targeted scenario. NbuGAN is experimentally validated: with 100 clean high-resolution images, NbuGAN creates 4275 high-resolution adversarial images that deceive 12 classification models trained on ImageNet for several clean-target combinations and expectations. Its average success rate is up to $93.3\percnt$93.3%, each high-resolution adversarial image being obtained in less than a minute on average. NbuGAN is compared to nine state-of-the-art black-box and white-box attacks. NbuGAN not only significantly outperforms the black-box attacks, but its remarkable speed, its success rates and the exceptional visual quality of the created high-resolution adversarial images, make NbuGAN highly competitive, both intrinsically and comparatively, even against white-box attacks.
Accurate forecasting of non-stationary time series, such as pandemic data, is particularly challenging due to the presence of concept drifts, i.e. changes in the underlying data-generating process over time. Conventional batch learning models often suffer from performance degradation in these settings, as they are trained on past data distributions that may no longer reflect current conditions. While incremental learning enables continuous model updates to better track such changes, it is often regarded as an approximation of fully retrained batch models. In this work, we introduce EFRT-DD (Extremely Fast Regression Tree with Drift Detection) and CDR (Collaborative Drift-Driven Regression), two contributions for adaptive stream regression. EFRT-DD is an eager incremental regression tree that continuously revisits its internal structure to remain aligned with evolving data, while CDR is a collaborative framework that integrates incremental and batch learners to reconcile the drift-adaptation capabilities of incremental models with the predictive strength of batch learners. Using the COVID-19 pandemic as a representative and highly non-stationary real-world benchmark, and under a strict prequential evaluation protocol, our experiments show that EFRT-DD improves upon state-of-the-art incremental regression trees, while the CDR framework further enhances predictive performance compared to standalone incremental or batch models.
In recent decades, indoor positioning systems have progressed rapidly to service various areas, including navigation systems, tracking devices, and autonomous vehicle management. In contrast to the outdoor positioning system, where the global navigation satellite system is a known typical technology, there is no standard indoor environment, despite the availability of miscellaneous connections of wireless networks, including a new lower-power version of Bluetooth, Radiofrequency (RF), Wi-Fi, RFID, vlc, and Ultrasonic. Within these technologies, the Ultrasonic position system (UPS) provides an outstanding implementation option because of its high precision and cheap cost. Although UPS provides several benefits, it faces obstacles from propagation circumstances like multipath and Doppler effect. Consequently, although various ultrasonic systems have been developed in recent decades, the issues remain an active and open subject. In this survey article, a comparison will be made between current scientific studies that will be categorized into several categories based on the method used to determine the location of objects as time-of-flight (ToA), round-trip-time (RTT), received-signal-strength (RSSI), and aoa. Furthermore, this study aims to assess the precision and efficacy of existing positioning systems comprehensively. The survey endeavours to present a thorough overview of critical attributes of alternative systems, encompassing their respective positioning methodologies, techniques, and measurement algorithms. Additionally, we will appraise these systems' constraints, considering factors such as precision, noise levels, and detection methodologies. Advanced simulation tools will be employed to replicate diverse scenarios to facilitate this evaluation.
Mining frequent weighted utility closed patterns (FWUCPs) from dynamic quantitative data streams is a challenging yet practically important task. The task is to discover meaningful patterns in real-time streams with dynamically changing item weights. To address this challenge, we formally define the problem of FWUCP mining under a sliding window model with dynamically changing weights. We propose two novel single-pass algorithms: CTC-MINER, based on a cyclic tidset (CTset) structure, and SAC-MINER, which exploits a stream-adaptive tree (SA-Tree) with compressed SAN-List. Both algorithms are designed to efficiently handle continuous updates while ensuring correct closure checking in a streaming setting. Extensive experiments on five benchmark datasets demonstrate the effectiveness of the proposed framework. The results show that SAC-MINER achieves significant advantages in runtime and scalability on dense and moderately sparse datasets, while CTC-MINER remains competitive on extremely sparse or low-cardinality datasets. These findings highlight the trade-offs between compression-based and tidset-based strategies and confirm the robustness of our framework for real-world, large-scale data stream mining.
Purpose: Breast cancer is a major global health concern, with early detection being critical for effective treatment and reduced mortality. While ultrasound imaging is a widely used diagnostic tool, manual interpretation is time-consuming and subject to variability among clinicians. Automated, accurate, and interpretable approaches for segmentation and classification are therefore essential to support clinical decision-making. Methods: This study introduces XAI-CAD (Explainable Artificial Intelligence-based Computer-Aided Diagnosis), a unified and interpretable multitask deep learning framework tailored for breast ultrasound imaging. XAI-CAD combines a U-Net-based segmentation backbone with pretrained encoders for classification, performing lesion segmentation and pathological diagnosis simultaneously. To enhance interpretability, the framework incorporates Grad-CAM and attention-based visualization techniques, highlighting regions most relevant to the classification decision. The system was evaluated on the publicly available Breast Ultrasound Images (BUSI) dataset, with rigorous statistical validation to ensure reliability. Results: The proposed framework achieved over 90% Dice coefficient, precision, and recall for segmentation, and outperformed several recent state-of-the-art models in classification accuracy. Explainability analyzes demonstrated that the model consistently focuses on clinically significant areas, providing transparent decision support. Conclusion: XAI-CAD offers a robust, accurate, and interpretable solution for breast cancer detection in ultrasound images. Its multitask and explainable design presents a clear methodological contribution and can be adapted to other medical imaging modalities, supporting both improved diagnostic performance and clinician trust.
Wireless communication technology faces challenges in terms of security and efficiency, and there is an urgent need to adopt intelligent methods to counter interference and improve the wireless communication experience in different interference environments. Therefore, a resistance-to-interference technology has been proposed to address the communication problems faced by users in different scenarios. Experimental results show that the multi-user distributed interference resistance alliance formation (MDACF) algorithm, compared with direct transmission, has a transmission efficiency increase of 12%. The interference resistance algorithm based on offline UCB1 adjustment of Q-learning (DUCBQ) can find the path with the minimum hop count in approximately 25 iterations, and the proximal policy optimization (PPO) algorithm, after training, has an average data packet reception rate of approximately 0.91. The average performance estimation value of the intelligent interference resistance technology proposed in the study is 0.8435. The three intelligent interference resistance algorithms proposed in the study have feasibility and interference resistance capabilities when facing interference environments.
With the increase of mission complexity in urban environments, distributed unmanned aerial vehicle network puts higher requirements on communication connectivity and perception accuracy. Therefore, a topology optimization method for integrated sensing and communication distributed unmanned aerial vehicle network based on orthogonal frequency division multiplexing signals is constructed in the study. The method firstly constructs orthogonal frequency division multiplexing signal model and fuses delay-Doppler domain characteristics for target sensing. Meanwhile, the integrated sensing and communication distributed optimization algorithm is designed to achieve the joint enhancement of sensing coverage rate and communication connectivity rate through a distributed iterative strategy. The experimental results indicated that the proposed method could reach 0.98 in perception accuracy. In terms of communication connectivity rate, it could reach up to 0.98 in high-density deployment, and could maintain above 0.82 in low-density deployment. In terms of sensing coverage rate, the average value reached 94.5%. Simulation results indicated that the proposed framework achieved effective topology awareness and connectivity preservation under sensing-assisted optimization. The reported perception-related performance reflects network-level topology cognition and information awareness, rather than physical-layer sensing accuracy. The results emphasize the potential benefits of incorporating OFDM-inspired sensing indicators into the optimization of distributed UAV topologies for dynamic deployment scenarios.
Autonomous vehicles have become increasingly popular from the last few years because of their numerous benefits, such as minimum travel time, enhanced safety, and improved fuel economy. Many vehicle manufacturing companies and tech giants are working on this technology to make fully autonomous vehicles or strengthen their existing driver-less vehicles. These vehicles use complex, advanced, and sophisticated hardware technologies. However, the software is an equally important feature because it must operate all functions seamlessly in sync with other vehicle components. The software must analyze large volumes of data to make quick real-time decisions, so any vulnerabilities or bugs can be a severe problem to the vehicle and the passengers riding in it. Many researchers have proposed various software defect prediction schemes for different projects and applications, but most of them have focussed on specific software issues and excluded others. Thus, their methods cannot be applied to the software of autonomous vehicles. In this paper, we propose an improved Artificial Neural Network (ANN) model, called Dropout-Artificial Neural Network (D-ANN), to solve this problem of defect prediction in autonomous vehicles. This inclusive model can consider all the parameters simultaneously for effective bug prediction. The proposed model can be used for the software of any autonomous vehicles, and it is trained and evaluated using standard methods. The results obtained show that the proposed model predicts software defects with higher accuracy than other models.
This paper addresses the problem of ensuring high-capacity, reliable wireless connectivity in temporary or disaster-stricken urban environments. We propose an optical intelligent reflecting surface (OIRS)-enhanced free-space optical (FSO) backhaul framework integrated with unmanned aerial vehicles (UAVs) to overcome the limitations of conventional FSO systems, such as line-of-sight (LOS) blockages. Specifically, OIRS deployed on high-rise buildings intelligently redirects optical beams, creating virtual LOS paths and stabilizing FSO links between the ground station and UAVs. To fully exploit this architecture, we develop a comprehensive channel model that accounts for path loss, turbulence-induced fading, and pointing errors, and introduce the OPT-UAV algorithm for joint UAVs 3D placement, user association, and bandwidth allocation under backhaul capacity constraints. Numerical results demonstrate that the proposed framework improves fairness and coverage by serving up to 67% of mobile users, compared to 62% with baseline methods, such as the TLA algorithm, while maintaining competitive throughput. These results confirm the potential of OIRS-assisted mixed FSO/RF systems as a scalable and resilient solution for smart city deployments, disaster recovery, and temporary high-traffic events.
At present, people's livelihood demand data collection is often affected by factors such as limited geographical coverage, a high proportion of sudden demands, and inconsistent data recording periods. The scarcity of small samples and non-stationary distribution are common. To this end, this study proposes a people's livelihood demand prediction strategy that integrates improved deep convolutional generative adversarial networks and reinforcement learning. Firstly, the convolutional layers and batch normalization structure of the adversarial network are generated through deep convolution. Secondly, a reinforcement learning architecture is constructed, in which the problem representation module adopts a bidirectional long short-term memory network combined with an attention mechanism. The results show that the model has the best prediction accuracy for long-term trends, and the symmetric mean absolute percentage errors are less than 10%. The short-term root mean square error of dataset A is less than 1.05. In the response test, the 30-3 window is optimal for traffic type, and the 30-7 window is optimal for industrial noise in the long term. The proposed model can effectively improve the robustness of demand prediction and provide technical support for the refined management of government affairs.
Intrusion detection in Internet of Things (IoT) environments presents challenges due to the diversity of connected devices and their resource limitations. IoT networks generate complex, imbalanced traffic where benign activity predominates over attack instances. This imbalance hampers the performance of traditional intrusion detection systems, which struggle to generalize effectively. In this study, we present a deep neural network-based system that leverages advanced data balancing techniques - such as subsampling, Synthetic Minority Over-sampling Technique (SMOTE), and Tomek Links - combined with cross-validation to enhance the model's generalization and minimize overfitting. Evaluations on CICIDS2017, UNSW-NB15, and BoT-IoT datasets showed accuracy rates of 99.2%, 99.7%, and 99.8%, respectively. These results demonstrate that our methodology outperforms traditional models, especially in detecting minority attack classes, which were previously challenging due to data imbalance. The use of data balancing and cross-validation significantly improved model stability and sensitivity to diverse attack scenarios. Our findings suggest that incorporating these techniques can substantially enhance the security of IoT environments, providing a robust approach for differentiating between normal and malicious activities, thus contributing to more reliable and scalable intrusion detection systems.
Traditional channel estimation and coding control methods suffer from low estimation accuracy and slow response, making it difficult to meet the demands for efficient and stable communication. Therefore, a channel estimation method, based on meta-learning and a Back Propagation Neural Network, is proposed. An adaptive coding dynamic control algorithm, which combines Reinforcement Learning and Adaptive Modulation and Coding linkage, is developed to achieve fast and accurate channel state estimation and real-time dynamic adjustment of coding strategies. Experimental results show that, in terms of channel judgment accuracy, the proposed algorithm achieves up to 99%, outperforming the best performance of the comparison algorithm, which is 97%. The average spectral efficiency reaches 6.4 bps/Hz, significantly higher than the 4.2 bps/Hz of the comparison algorithm. Moreover, the spectral efficiency fluctuation range of the proposed algorithm is from 5.9 bps/Hz to 8 bps/Hz, showing better performance compared to the 3 bps/Hz minimum fluctuation range of the comparison algorithm. These results demonstrate that the proposed algorithm has significant advantages in channel estimation accuracy and coding control performance. It provides a new approach for efficient and stable transmission in satellite laser communication and contributes to the advancement of intelligent and efficient technology in this field.
Dynamic network communities evolve over time. Traditional methods suffer from declining partition quality and trajectory discontinuities, compromising the decision accuracy and real-time performance of critical applications such as social recommendation and anomaly detection. Addressing this issue enhances analysis of complex systems' dynamic behaviours. This study proposes a multi-objective dynamic community detection algorithm for complex networks, with an adaptive mutation strategy. It optimizes semantic consistency and topological coherence across time slice communities through dynamic mutation parameter adaptive adjustment and hierarchical event detection. Experimental results demonstrate that the proposed algorithm performs well on both simulated and real dynamic networks: peak community partitioning quality reaches 0.91, the average community cohesion (modularity) reaches 0.87, the average temporal smoothness reaches 0.84, and the peak community event detection rate reaches 94.8%, outperforming baselines; Additionally, the burst response latency is below 90 ms, with significant improvements in computational efficiency and memory usage. The algorithm demonstrates long-term stability (efficiency decay rate <8%), with community overlap detection rates above 0.80. The dynamic community detection platform developed through research can improve the accuracy and real-time processing capability of analyzing complex network evolution patterns, providing efficient and scalable solutions for dynamic decision-making in scenarios such as social recommendation and anomaly monitoring.
