Abstract Detecting malignancy in pulmonary nodules holds significant clinical importance, yet existing image classification methods often struggle with inadequate feature integration and ineffective loss functions. This study proposes two innovative strategies to address these limitations: first, we introduce a multiscale feature weighted fusion technique that enhances the integration of features across different scales, allowing the model to prioritize critical pixel locations essential for accurate diagnosis. Second, we combine contrastive loss with binary cross-entropy within our training framework to improve learning from both similarities and differences among paired samples, which fosters better discrimination between similar nodules while maintaining sensitivity to variations across classes. Besides, our proposed methodologies demonstrate promising performance improvements in detecting pulmonary nodule malignancy, leading to enhanced performance and reliability compared to conventional approaches.

Abstract This study analyzes the resilience of open source smart home platforms, namely, Home Assistant, RaspberryMatic, HomeBridge, Nymea, and OpenHABian, against distributed denial of service (DDoS) attacks such as TCP SYN flood, UDP flood, and Internet Control Message Protocol (ICMP) flood in IPv4 and IPv6 networks. As the IoT ecosystem grows, so does the importance of cybersecurity for smart home platforms. The research evaluates the impact of different attack intensities on the availability and stability of the platforms, comparing their performance in both network protocols. Experimental results show differences in the resilience of each platform. IPv6 showed higher resilience to high frequency DDoS attacks, while IPv4 showed higher stability at moderate load levels. The results highlight the need to optimize network protocols and security mechanisms to increase the reliability and resilience of smart homes to DDoS attacks.

Abstract The nonconvex second-order cone (nonconvex SOC) is a nonconvex extension to the convex second-order cone, in the sense that it consists of any vector divided into two sub-vectors for which the Euclidean norm of the first sub-vector is at least as large as the Euclidean norm of the second sub-vector. This cone can be used to reformulate nonconvex quadratic programs in conic format and can arise in real-world applications. In an attempt to obtain an approximate solution for optimization problems over the nonconvex SOC, in this article, we use a heuristic algorithm based on the alternating direction method of multipliers to solve them, which is the core result of our study. More specifically, the approach is built in two steps: a convex optimization problem comes first, followed by a nonconvex conic optimization. The problem in the second phase can lead to an inexact solution. Our strategy makes use of an approximate projection onto the nonconvex cone. The question of convergence remains open.

Abstract Considering the shortcomings of traditional video multi-target recognition technology in rapidly identifying criminal suspects in complex scenes, a multi-target recognition optimization method based on video tracking technology is proposed. This method constructs a multi-target recognition algorithm based on video feature matching, and introduces the Kalman filter algorithm to improve the accuracy and real-time recognition of criminal suspects through the definition of feature vector and similarity function. Experiments showed that the model proposed in the study performed exceptionally well in terms of tracking error; the highest precision was 94.75%, the recall rate was 96.59%, the tracking error of the horizontal axis was only 3.75%, and the tracking error of the vertical axis was 3.27%. In the crime detection video application, the accuracy–recall curve of the model was 0.94, and the feature recall rate was 94.83%, verifying the effectiveness and robustness of the model in complex and fast scenes. The results show that the proposed model has good feasibility and robustness in rapidly identifying criminal suspects. In addition, the work offered new technical concepts for improving target tracking precision and adapting to real-time scene changes, opening new research avenues in the field of multi-target recognition.

Abstract In web service composition, the Quality of Service (QoS) prediction applications based on the statistical point estimation method in accuracy consist of many challenges. Aiming at allowing users to select the web service composition based on their requirements, this study proposed a method based on Bootstrap to estimate and predict the QoS confidence interval for web service composition (QCI-WSC). The QCI-WSC first indicates the structure of the web service composition and simplifies the structure model. Apart from that, the QoS estimation interval can be calculated by the historical QoS data, which are invoked by users. Meanwhile, the user similarity is calculated, and the QoS of web service invoked by the similar users is used to predict QCI-WSC. Finally, the results of user-invoked web service composition QoS are verified by the average interval coverage rate, compared to the actual QoS values and prediction values of the other methods, such as adaptive QoS prediction method based on collaborative filtering (QACF) and QoS-Aware web service recommendation (WSRec). Additionally, in this work, dataset1 in WSDream is adopted to estimate and predict the QCI-WSC. Experiments show that the QoS confidence interval estimation results conform to the exponential distribution, and the validity of the QCI-WSC is proved. Furthermore, the average interval probability of the prediction algorithm was more than 75%. The QCI-WSC can accurately cover the actual QoS values of the web service composition and most of the accurate QoS values predicted by QACF and WSRec. It effectively improves the selectivity of service, which provides web service composition featuring better quality for users.

Abstract Regulating traffic in cities plays a crucial role in addressing climate change. The rapid advancement of artificial intelligence technologies offers new opportunities for managing urban traffic within the framework of Smart Cities. One common method for classifying traffic is the level of service (LoS) criterion, which evaluates traffic quality based on factors like density, speed, and location-specific characteristics. Because of these variations, LoS must be assessed individually for each location, often with expert assistance. In this article, we propose and compare several approaches for LoS classification using neural networks, fuzzy sets, and high-dimensional random vectors. Our goal is to reduce the reliance of LoS determination on local conditions, making the methodology adaptable across different locations. The results show that all three used methods achieved sufficient accuracy, which supports their potential integration with meteorological and pollution data for further applications.

Abstract In response to the current problems in teaching aesthetic education, such as a single teaching method and lack of innovation, the study uses digital media technology to improve it. First, Maya software is used to construct a three-dimensional (3D) model of ancient architecture, and then the UNity3D engine is used to create a roaming scene for students to experience the heritage and elegance of ancient architecture in the aesthetic education classroom. To enable students to achieve an immersive teaching experience, the study uses Kinect somatosensory technology to enable students to interact with the simulated scenes. The students were first tracked by a combination of random forest algorithm and mean shift algorithm, and then the dynamic time warping algorithm was used for dynamic gesture recognition to ensure that students could move forward and backward through the gestures while using the simulation to navigate through the building. Through the above operations, the study completes the design of a virtual recreation system for ancient cultural buildings based on digital media technology. Through the experiment, it was obtained that 94.81% of the total number of students surveyed believed that this method could stimulate interest in learning about aesthetic education classes. The system designed by the research provides new ideas for the reform of the teaching mode of aesthetic education in universities.

Abstract The Internet of Things (IoT) encompasses a network of interconnected devices that collect, analyze, and exchange vast amounts of data. However, this connectivity creates opportunities for various types of cyberattacks, making IoT systems vulnerable and potentially leading to the compromise of sensitive information. Therefore, developing effective intrusion detection system (IDS) is one of the key challenges in IoT network security. The aim of this study is to develop a machine learning (ML) model for network traffic classification and attack detection in IoT environments. Through a comparative analysis of different algorithms, the study seeks to identify the model with the best performance, which could serve as a foundation for efficient IDS solutions tailored to the specific characteristics of IoT networks. The RT-IoT2022 dataset was used for experimental analysis, providing realistic framework for testing ML models, including k-nearest neighbors, Random Forest, XGBoost, multilayer perceptron, and various 1D convolutional neural network architectures. The study examines preprocessing techniques, focusing on dimensionality reduction (principal component analysis, variance inflation factor, Pearson’s test), outlier detection (interquartile range, Z-score, Isolation Forest), and transformation methods (Box–Cox, RobustScaler, Winsorization). Based on the results of the experiment, the most effective model and preprocessing technique were proposed.

Abstract This study aims to assess the effectiveness of various algorithms in the realms of supervised and semi-supervised learning applied to three multiclass facial image datasets: JAFFE, Georgia tech, and Yale. The datasets were partitioned into proportions of 80:20, 75:25, and 50:50 for supervised learning, while semi-supervised learning was conducted with labelled and unlabeled data ratios of 20:80, 25:75, and 50:50. The evaluated algorithms include convolutional neural networks (CNNs), decision tree, long short-term memory, K-nearest neighbors (K-NNs), multilayer perceptron, and support vector classifier (SVC), each with varying parameters. Experimental outcomes reveal that the performance of models depends on the dataset partitioning strategies and the type of algorithms used. Specifically, linear and polynomial SVC consistently yield favorable results in supervised learning, particularly demonstrating efficacy on the Georgia tech dataset. Conversely, on the JAFFE and Yale dataset, linear SVC and K-NN emerge as optimal choices. The inclusion of semi-supervised learning enhances insights, particularly evident in the Georgia tech dataset, where the combination of labeled and unlabeled data significantly improves accuracy, especially when leveraging linear SVC algorithm. Although there are some instances of sub-optimal performance in certain algorithms like CNN on specific datasets, this research provides comprehensive insights into the effectiveness of various models in contexts of limited-label learning. The implications of these findings are crucial in advancing the development of adaptive and robust facial recognition systems, especially in navigating datasets characterized by diverse variations and complexities.

Abstract Electric vehicles (EVs) are emerging as major energy consumers, offering numerous environmental and operational advantages such as reduced greenhouse gas emissions and lower reliance on fossil fuels. As the adoption of EVs accelerates globally, accurate forecasting of EV charging demand becomes increasingly critical for maximizing the efficiency, reliability, and profitability of charging infrastructure. However, many existing forecasting models fall short by neglecting the complex and dynamic influence of external factors – particularly weather conditions and calendar variables – which can significantly affect usage patterns. This study presents a robust forecasting framework that integrates historical charging data with both temporal and meteorological information to comprehensively evaluate their individual and combined impacts on EV charging behavior. Leveraging long short-term memory networks – effective in modeling time-series data – we evaluate the impact of contextual features on forecasting performance. Results show that calendar information notably improves accuracy, surpassing the effect of weather data. These insights help EV station operators optimize scheduling, reduce uncertainty in day-ahead energy planning, and support sustainability and grid stability.