Frontier Computing on Industrial Applications Volume 4

Deep learning has made significant improvement in natural language processing. Nowadays virtual assistants or chatbots attract attention of many researchers and are expected to be applied in more and more areas. We had designed and implemented an extensible financial virtual assistant using Genie framework. A new device (or skill) is developed to offer financial services in backend server cloud. The device and supported APIs (Application Programming Interface) are registered in an open repository Thingpedia. When Genie receives user utterances, it translates them into ThingTalk programs using a large deep-learning neural networks. Then, Genie executes the ThingTalk programs, which may invoke the financial services through the registered APIs. ThingTalk is a declarative programming language. Domain experts can easily describe financial services in high-level viewpoint with minimal knowledge and experiences of computer programming and system development, while complex services are implemented in backend servers and access through APIs. As a result, domain experts and computer engineers together can fast and easily build a virtual assistant that support natural language interface.

PET/CT is the preferred device for lung cancer and lymph node metastasis diagnosis, and mining effective features from PET/CT images to identify lung lymph node metastasis has important research significance and application value. Multi-phase PET/CT has temporal properties that can better represent changes in lesions’ structural and metabolic properties. Early-phase PET images can show a wide range of lesion areas. Delayed-phase PET images can show the high uptake properties of 18F-FDG in malignant tumor cells. Thus, multi-phase PET represents the variability of benign/malignant lesions better in the temporal dimension. This paper first proposes a metabolic enhancement method for lung lymph nodes and their microenvironment, a lymph node metastasis recognition network (LNMER-Net). The network has three branches: multi-modal early-phase feature fusion channel, multi-modal delayed-phase feature fusion channel, and single-modal metabolic decay channel. To enhance the feature of the lymph node region, a multi-receptive field-based feature extraction and feature space optimization (MRFO) method is proposed to extract lymph node features by multi-scale convolution operations and embed them in the multi-modal fusion channel. To exploit the information on the metabolic changes of the lesion in the early-phase and delayed-phase, differential results of the multi-phase PET images are fed into the single-modal metabolic decay channel to enhance the microenvironmental features. To verify its effectiveness, a multi-phase PET/CT dataset from China Medical University is used. The proposed method achieves 84.5%/82.9% in Accuracy/Recall, which is better than SOTA methods such as Res2Net, Comformer, and NextViT.

YouTube has become a major advertisement media for industries. People watch TV shows, movie trailers, news and episodes, on-line shows streaming as well. When people want to purchase something, they can find some unboxing reviews, experiences shares, using instructions and services introductions on YouTube before making orders. Traveling guides and educational tutorials are also can be found on YouTube. Thousands of YouTube channels need channel’s subscribers directly or indirectly affect the revenue of channel owners. In addition to YouTube’s own profit from watching, the subscribers are also an important reference when manufacturers want to choose key opinion leaders (KOL) or internet celebrities to promote their products or services. Since, maintaining a stable growth of channel subscribers and reducing the occurrence of unsubscribing are the key issue that every YouTuber needs to understand urgently. In this study, we used a survey to understand subscribers and YouTubers on unsubscribing, and analyzed some key factors for unsubscribing. The data analyses include descriptive analysis, factor analysis, reliability analysis, t-test, one-way ANOVA, and Pearson’s correlation analysis. Through this study, people can understand key factors for unsubscribing reasons on YouTube platform and the priority of these factors. Whether there is special unsubscribe factors for different channel types and attributes. The key contribution of this study is: a questionnaire on key factors for unsubscribing on the YouTube platform was designed, and it could be utilized for further use in related studies in the future. We also summarize the responses and suggestions of we-media channel owners on the factors for unsubscribing.

This project aims to study learners’ visual focus in online learning and their behavior patterns when engaging with different learning media and cognitive processes. It uses AI models and data analysis methods to evaluate learners’ online preferences and information processing.The project consists of three stages:(1) Collecting and Integration: Use visual movement devices to collect learners’ visual movement data and other data to analyze visitors’ visual behavior and identify key eye movement factors.(2) AI Applications: Use AI data mining technologies to identify the visual focus of the target learners through machine learning and data analysis.(3) Integrated Analyses of Online Learning: Optimize online learning design by combining visual focus analysis and information collected in visual mode. Collaborate with companies to establish an “adoption model for visual focus analysis of online education,” which will be verified by the online teaching platform.

We propose a fair distributed computing platform based on Distributed Ledger Technology (DLT) and performance measurements. The platform integrates DLT and federated learning, enabling users to train machine learning models on their local devices without compromising their privacy by sharing their data with a central server. Instead, only the trained model weights are uploaded to a central server for aggregation. To address privacy concerns associated with federated learning, we integrate various privacy-preserving methods, such as differential privacy, model pruning, and homomorphic encryption, into the platform framework. These techniques help protect user privacy while improving model accuracy. To address the non-IID data problem in federated learning, we use performance measurements to balance the training workload among users, and blacklist malicious users while incentivizing participation. DLT ensures the security and integrity of the platform by validating and recording all data transactions on the ledger. Overall, the proposed platform has the potential to revolutionize machine learning model training by making it more efficient, secure, fair, and transparent.

Studies on eye tracking have relied on wearable eye trackers and chin-resting eye trackers, but the high cost of equipment and the need to wear devices during experiments can lead to less natural facial movement. This study collected eye-tracking data by using a raw-video camera without adjusting any parameters. Python was used as the primary programming language. Eye tracking was adjusted through calculations of facial distance, and multitask cascaded convolutional networks were used to collect eye-tracking data. The corrected results were visualized, and linear regression was used to determine correction error. The root mean square error was 221.66, and the mean squared error was 260.48.

In natural language processing, the quality of language models impacts applications such as machine translation and speech recognition. GPT-2, a powerful auto-regressive model with 150 million parameters, performs exceptionally well in various tasks but struggles with computational efficiency for long sequences. We have developed an optimization strategy to mitigate this issue by randomly shortening the auto-regressive length during generation. Our strategy was tested on the GPT-2 medium model using BLEU as the evaluation metric. The results revealed significant improvements in the BLEU scores, with the optimized model outperforming the original. Furthermore, the optimization also improved scores in both the top and bottom 10% of the data. Despite the promising results, there is still room for further exploration and improvement. We are currently investigating adaptive adjustments to the auto-regressive length and applying this strategy to other models, such as GPT-3. In summary, our research proposes a new strategy that enhances GPT-2’s efficiency and boosts its performance, as evidenced by the improved BLEU scores. This strategy provides valuable insights for future language model optimization, holding the potential to advance the field of NLP.

This experiment classifies cases based on the top 20 most frequently used categories of case reasons found in civil summary court judgments provided by the Judicial Yuan of Taiwan from 2012 to 2022. We built case classifiers using two methods: machine learning with TF-IDF+SVM and deep learning with BERT. We then compared the results of both classifiers. In the classification results using TF-IDF+SVM, an accuracy of 89.3% was achieved, while with BERT, an accuracy of 93.825% was achieved.

Virtual reality (VR) games have gained significant popularity in recent years, offering immersive and interactive experiences. The success of VR games relies heavily on innovative interaction modes that enhance player engagement and immersion. This paper explores the concept of innovative interaction modes in VR games and their impact on player experiences. We examine various forms of interaction modes, including gesture-based controls, motion tracking, and haptic feedback, and analyze their effectiveness in enhancing gameplay. Furthermore, we discuss the challenges and opportunities in designing and implementing innovative interaction modes. Through a comprehensive review of existing research and case studies, this paper provides insights into the potential of innovative interaction modes to revolutionize the gaming industry. The findings underscore the importance of continuous innovation and experimentation to create compelling and immersive VR game experiences. By understanding and leveraging these innovative interaction modes, game developers can deliver more engaging and memorable gameplay, transforming the way players interact with virtual worlds. In this study, we conducted a prototype system to evaluate the impact of innovative interaction modes on player engagement and gameplay experiences in VR games. We developed two different interaction modes: gesture-based controls and in-game control tool.

Both blended learning and work examples are used to enhance student learning. Due to the temporary suspension of classes due to COVID-19, the face-to-face course was temporarily converted to a full-scale online course. In order to provide a consistent learning foundation for students of all levels, The Next Generation Art course is based on segmented worked examples, supplemented by tutorial videos and supplementary videos. At the end of the semester, a questionnaire based on the TAM scale was administered to three classes in two school districts. Gender, 3D ability, and basic information of topic type were analyzed by Independent Sample t-test. The results showed that females had better 3D ability than males, which was also reflected in the TAM, and females preferred and were willing to apply the technology of the course. As expected, those with lower 3D ability chose 2D format for their topics. The relationship between the number of professional credits and 3D ability was analyzed by ANOVA. Unlike the prediction, it is not the case that more courses equals higher ability, but rather the students with moderate number of 3D courses considered themselves to have the highest 3D ability. It is possible that the Dunning-Kruger effect is caused by the worked example and the blended learning.

This study explores the application of artificial intelligence in lathe cutting machine tools in smart manufacturing. Long-term processing will cause thermal deformation of the lathe cutting tool machine, which will cause displacement errors of the cutting head and damage to the final product. Using time-series thermal compensation, the research develops a predictive system that can be applied in industry using edge computing technology to predict the thermal displacement of machine tools. The study conducted two experiments to optimize the temperature prediction model and predict the five-axis displacement of the temperature point. Furthermore, a genetic algorithm is used to optimize the LSTM model to predict the thermal displacement of the machine tool. The results show that the GA-LSTM model achieved a thermal displacement prediction accuracy of 0.99, while the average accuracy of the LSTM, GRU, and XGBoost models was 0.97. Based on the analysis of training time and model accuracy, the study recommends using LSTM, GRU, and XGBoost models to design and apply to systems that use edge devices such as Raspberry Pi for thermal compensation.

The computing power and storage requirements of the Internet of Things (IoT) are likely to increase substantially in the future years. Because of the rapid development of both machine learning (ML) and the Internet of Things (IoT), vast volumes of data created by edge devices such as smartphones, laptops, and artificial intelligence (AI) speakers have been widely used to train ML models. In this study, we used a cluster-based Blockchain method in the Multi-Access Edge Computing (MEC, also known as Mobile Edge Computing) for markets and technological services. We describe a generalized stochastic block model (SBM) for edge computing applications based on the proposed taxonomy. These mobile edge wireless devices (WD) provide efficient resource allocation in mobile network situations. In our studies, we compared the approximate solutions obtained by the SBM to those generated by the cluster-based Blockchain algorithm. However, the high latency and low scalability of traditional blockchain systems limit mobile transactions on the public blockchain. To reduce the consumption of competitive mobile transactions created by linear sequencing blocks, reconstructed blockchain systems have been developed. This study’s use of cluster-based blockchain systems provides speedy confirmation and great scalability without significantly compromising security.

Location-based advertising (LBA) has been popular for several years, and the amount of global investment is increasing year by year. Nowadays, in the vigorous development of vehicle vision systems, many recognition tasks can be completed by combining You Only Look Once version 7 (Yolo-v7) object detection algorithms to apply automotive applications, and also involve a QR codes decoding method with deblurring generative adversarial network version 2(DeblurGAN-v2), which can capture the QR codes set on the route in real-time to obtain the LBA placed by the merchant, the results show that the proposed method outperforms the other object detection model and deblurring model, it obtains more efficient for scanning QR codes.

Histology image annotation is costly and time-consuming. Utilizing Positive and Unlabeled (PU) data for model training offers a more resource-efficient alternative. However, previous methods for PU learning suffer from the noise arising from label assignment to unlabeled data. We observe that predictions on noisy data lack consistency under data augmentation. In this paper, we present Field of View (FoV) consistency regularization for PU segmentation in histology images, which effectively reduces the noise influence by promoting consistent predictions across varying FoVs. Using only 20% of positive labels on the Glas Dataset, our approach outperforms previous methods, achieving a Dice score of 90.69%-almost reaching the fully supervised result of 93.30%. Source code is available at: https://github.com/lzaya/PU_with_FoV .

Attribute-based encryption is one of the most suitable access control mechanism for modern data sharing models. To provide better performance, lots of attribute-based encryption schemes are constructed over without pairings. However, these schemes are either with no security proofs or broken. In this manuscript, we give the cryptanalysis of two key-policy attribute-based encryption schemes for weighted threshold gates. We propose two attack methods, the first one is able to generate valid private keys without the master secret keys, and the second one is able to recover the master secret key when an attacker gathers enough number of private keys. Moreover, an improved schemes is given in this manuscript. We also present a security analysis to show that our improved scheme fix the security flaws with only one pairing added.

This paper explores the use of virtual reality in developing educational materials for the conservation of energy law in the field of natural sciences, specifically focusing on gravitational potential energy and elastic potential energy. Through immersive experiences in virtual reality, students are provided with an enjoyable learning opportunity. The visual experience in virtual reality is designed to simulate scenarios involving three different gravitational fields of planets, allowing learners to break free from the constraints of reality and experience the conversion between potential energy and kinetic energy within the context of energy conservation. Students are immersed in an engaging learning environment, where they can truly grasp the essence of Newtonian mechanics.

Spatial keyword queries find extensive applications in geographic information systems like Facebook and Instagram. The collective spatial keyword query (CSKQ) plays a crucial role among the various types of queries. This query aims to retrieve a set of Points of Interest (POIs) that collectively cover the specified keywords while being in proximity to both the query location and other objects. To evaluate the spatial cost of a set of POIs in CSKQ, we introduce the Edge-Based Collective Nine-Area Tree Index (EBCNA). By incorporating edge information and POIs into the NA-tree structure, the EBCNA offers a comprehensive solution. All edge information, including POIs, is stored in the leaf nodes, and each edge links to its adjacent edges via pointers. This design enables direct retrieval of edge information and POIs without repeatedly trailing back to the root node. Through a comparative analysis, we have demonstrated our proposed method’s superior performance compared to the existing one.

This study explores the use of Generative Adversarial Networks (GANs) to generate wafer-level Wafer Acceptance Test (WAT) and Chip Probe (CP) test data in semiconductor manufacturing processes, and their application in relevant process and Design-Technology Co-Optimization (DTCO). The generated virtual silicon data includes device performance, physical-electrical characteristics, distribution of wafer process parameters, and implicit information on wafer-level features such as uniformity and defects. This approach enables interdisciplinary teams to overcome data acquisition barriers while ensuring data confidentiality, and it holds significant potential for the development of advanced Electronic Design Automation (EDA) tools in co-optimizing process and chip design flows.

Automated recognition of pneumonia from chest CT plays an important role in the subsequent clinical treatment for patients. While a few pioneering works only focus on several random slices from chest CT image, thus they have ignored the anatomical dependency information of local lesions. Considering it, this paper explores a novel automatic classification method for pneumonia detection based on fusing regional and global information, which not only improves detection performance, but also provides explainable diagnostic basis for radiologists. Firstly, identifying the interested local region by a lesion detection module, then we extracts the correlation relationship between local regions through a graph attention module. The image-level classification results can be acquired by fusing the information of global and local region. To realize the detection of full CT sequence, a person-level classifier is designed in the proposed model. In the experiment, we collected 781 chest CT sequences in total corresponding to 274 cases of viral pneumonia patients, 285 cases of bacterial pneumonia patients and 222 cases of healthy people. The experimental results show that our model achieves the accuracy of 95.5%, with 95.6% precision and 0.991 AUC. The recall and F1 score are 95.8% and 95.7% respectively, which outperformed previous works. Therefore, our method can be regarded as an efficient assisted tool in the diagnose of pneumonia.

The integration of deep learning (DL) and digital breast tomosynthesis (DBT) presents a unique opportunity to improve the reliability of breast cancer (BC) detection and diagnosis while accommodating novel imaging techniques. This study utilizes the publicly available Mammographic Image Analysis Society (MIAS) database v1.21 to evaluate DL algorithms in identifying and categorizing cancerous tissue. The dataset has undergone preprocessing and has been confirmed to be of exceptional quality. Transfer learning techniques are employed with three pre-trained models - MobileNet, Xception, DenseNet, and MobileNet LSTM - to improve performance on the target task. Stacking ensemble learning techniques will be utilized to combine the predictions of the best-performing models to make the final prediction for the presence of BC. The evaluation will measure the performance of each model using standard evaluation metrics, including accuracy (ACC), precision (PREC), recall (REC), and F1-score (F1-S). This study highlights the potential of DL in enhancing diagnostic imaging and advancing healthcare.

Consider the high dynamics of traffic loading and resource provision on network hosts that forward data flows along a particular path between two endpoints. The Quick UDP Internet Connect (QUIC) protocol performs better than TCP for its effects in shortening the time of connection establishment and data transmission between two endpoints. Recent studies attempted to exploit the notion of multipath QUIC that forwards the data over multiple paths. Using the multipath QUIC can not only augment the total bandwidth capacity but also avoid traffic congestion on some paths. In this paper, our study proposes a novel multipath QUIC scheme which is able to minimize the flow completion time of multipath QUIC by jointly utilizing two measures of path delay and packet loss rate on a path. Experimental results show that the proposed algorithm is superior to other scheduling schemes, including naive QUIC and Lowest-RTT-First QUIC.

This research aims to develop an AI-based Ergonomics risk hazard posture recognition system to help reduce the risk of injury to workers and improve work safety in factories and warehouses. The background shows that ergonomic risk hazards are one of the most important risk factors in the workplace, among which the risk of posture hazards is higher when the human body is carrying objects. Otherwise, KIM-LHC (Key Indicator Methods - Lifting/Holding/Carrying) was used as the basis for posture determination, and the human posture information was converted into data by Movenet, and then build the neural network classification model used to recognize and analysis human pose, finally integrated into the app built by flutter. The app built by Flutter is finally integrated. In order to verify the performance of the system, it conducted experiments by actual video recording, and the results showed that the verification accuracy of the app could reach over 97%, and successfully identified the dangerous postures that might cause injury risks to workers, and the app was easy to understand and practical. In summary, this research developed an AI-based Ergonomics risk-hazard posture recognition system, which is important for improving workplace safety.

This research presents a new music generation model and a novel MIDI data format for MIDI music generation. This innovative data format allows us to process MIDI music in a manner analogous to video analysis. Initially, the model employs Convolutional Neural Networks (CNN) as an encoder to effectively capture local and global features within the musical data. Subsequently, we utilize a Transformer as a decoder, leveraging its self-attention mechanism to handle the long-term dependencies present in music data. In the training process, an interactive chaotic algorithm is introduced to update the model’s weights, assisting the model in avoiding entrapment in local optima. This enhances the learning efficiency of the model and improves the quality of the generated output, enabling the model to generate music, including accompaniment, that aligns with human aesthetics from any given melody.

Load balancing is an important issue in multi-tenant clouds to ensure the load balanced on computing resources belonged to different tenants. A containerized multi-tenant environment could be managed by Kubernetes using different scheduling policies. For improving scheduling performance of Kubernetes, Apache Yunikorn project provides the fine-grain control by hierarchical resource queues to enhance the resource utilization. However, the fairness-aware load balance among tenants is missed in Yunikorn, which may result in poor resource utilization in some tenants. This paper proposes a fairness-aware load balance policy for multi-tenant environments to keep the balance of resources allocated in different tenants, and also the balance of the utilizations of different resources in a computing node. Experimental results show the superiority of the proposed policy.

To comply with the specific requirements of smart grids, Chen et al. proposed a data aggregation scheme by utilizing double blockchains and the Paillier cryptosystem that is an additive homomorphic encryption system for public-key cryptography. Chen et al. claimed that their scheme could resist various attacks and ensure data confidentiality, data integrity, validity, identity anonymity and authenticity. However, after thoroughly analyzing their scheme, we find that it suffers from five flaws. Firstly, anonymity is not ensured as claimed. Secondly, private keys of smart meters and fog nodes can be easily retrieved. Thirdly, after a smart meter or a fog node’s private key is revealed, a malicious entity can impersonate it and generate a valid signature of the forged data’s ciphertext. Fourthly, in both of the UA-blockchain generation phase and FA-blockchain generation phase, the signature verification will never succeed. Fifthly, some statements in Chen et al.’s scheme are inaccurate or missing such that their scheme cannot work as claimed. The details of how these flaws damage Chen et al.’s scheme are shown in this paper.

In this study, transfer learning techniques will be used for model training, using edge computing [1] and deep learning object detection technology, combined with image road pothole detection applications, and deploying devices and tools that accelerate neural network operations, including DeepStream [2] and Intel NCS2. The performance and accuracy of model recognition will be compared, and finally, real-time streaming video technology will be used to present the results on the web. According to the experimental results, the best model achieved an mAP of 70.% in YOLOv4-tiny-3l, and in terms of operating efficiency, deployment on Jetson Xavier NX using DeepStream for acceleration can achieve 30FPS. Finally, the deep learning model recognizes the screen presented on the web. This application can improve the accuracy of Pavement Distress identification and help road maintenance units improve the efficiency of repairing roads.

Zebrafish (Danio rerio) is an ideal model organism for biological research due to its ease of breeding, maintenance, observation, and complete genome sequencing. As a small aquatic organism with a body length of about 3–5 cm, zebrafish mainly exhibits its behavior through swimming in water. Therefore, trajectory tracking is crucial for a deep understanding of zebrafish behavior and physiological states, as well as for revealing its associations with specific diseases. In addition, zebrafish is widely used for drug screening and toxicology testing to explore the underlying neural and physiological mechanisms. Because of the high efficiency, accuracy, structural simplicity, and versatility of YOLO series models in object detection, they have become one of the preferred deep learning models for many researchers and developers. In this study, a model trained using YOLOv7 was proposed to track the movement trajectories of zebrafish and classify their behaviors into three categories: swimming, sinking, and static, through time-series sorting. According to experimental testing, our method exhibits excellent performance in detecting zebrafish movement trajectories. On a test set consisting of one frame per second, the model achieved a 100% accuracy rate and a 100% recall rate, demonstrating its potential in automated trajectory tracking.

This study aims to address the communication barriers related to speech for individuals with cerebral palsy, with the goal of using technological methods to assist or alleviate difficulties in oral communication. To achieve this, the study plans to analyze and test mainstream speech recognition services or platforms available in the market to understand their current speech recognition capabilities for individuals with cerebral palsy, and explore the possibility of assisting them in solving their communication problems, in order to enhance their quality of life and promote their social skills. As the author is a person with congenital cerebral palsy, the study is particularly meaningful to him because the congenital brain damage affecting the nervous system has made his speech unclear, seriously affecting his ability to express himself orally. Therefore, the author plans to record a dataset of speech samples from individuals with cerebral palsy, collecting conversations from various aspects of daily life. This dataset will be tested and analyzed using mainstream speech recognition services such as Google, Microsoft, and YaTing, among others, in order to infer the current difficulties in speech recognition technology for individuals with cerebral palsy and propose potential solutions for oral communication barriers, with the hope that the contribution of this research will promote the development of mature assistive technologies for individuals with communication difficulties in the near future.

This study is based on deep learning techniques, which compare various detection algorithms and implement the suitable one for firework detection. The considered factors include streaming, speed, accuracy, and portability. Through a detection algorithm, it can simultaneously identify the positions of smoke and fire, providing subsequent control of fire or other applications. After comparison, we plan to perform detection results in a streaming manner, where only real-time detection of the captured scene is carried out. The system can notify people or teams in need of notification via the network.

With many research results and the development of related tools, user-level intrusion detection and prevention systems (IDPS) have been widely used to defend systems against network attacks. However, there are still bottlenecks in their high packet drop rate and low detection efficiency under heavy network traffic. In contrast, kernel-level IDPS has a higher packet detection rate and higher efficiency, whereas kernel-level design faces many challenges. The system designed with the monolithic architecture has high performance. The dynamically loadable module architecture design has higher flexibility and scalability; however, the increased operating costs lower system performance.This paper explores the modular architecture of kernel-level IDPS that can expand or reconfigure system functions through dynamic plug-in modules and maintain the system’s stability and high performance. We have developed a lightweight, high-efficiency, scalable, and highly modular kernel-level IDPS named mKIPS. This modular architecture divides the system into several kernel modules, in which functional components can be dynamically inserted or removed during runtime to adapt to changing demands. Therefore, administrators can control the IDPS’s packet processing by mounting modules of different versions and functions for their needs. Besides, mKIPS dispatches packets to various cores for processing through software and hardware functions by properly setting the IRQ affinity and using Receive Packet Steering technology. As a result, the load of each core can be more balanced to utilize the multicores. Experimental results show that our mKIPS can achieve a high detection rate and efficiency.

The COVID-19 pandemic has revitalized focus on predictive models, but scant research has been devoted to modeling game transmission, and current models are inadequate in this regard. To predict the spread of games within the population, this paper proposes the “addiction individuals”, a new group based on the three groups of the SIR model. We applied the SIAR model, designed based on differential equations, to predict game transmission within this population. The SIAR model was validated on an existing dataset and compared with the traditional SIR model, demonstrating its greater accuracy.

In this study, we propose a novel pre-learning approach for genetic programming (GP) that aims to investigate the effect of the probability of being selected for each operator. Furthermore, we present a technique that combines chaos theory and searches for a relatively good possibility mapping for each operator using one-dimensional chaotic mapping. We conducted several sets of comparative experiments on real-world data to test the viability of the proposal. These experiments included comparisons with conventional GP, examination of the impact of various chaotic mappings on the proposed algorithm, and implementation of different optimization strategies to find the relative optimal probability mapping. The experimental results demonstrate that the proposed method can achieve better results than conventional GP in the tested dataset, without considering the total quantitative calculation amount. Through statistical tests, it has been proven that the proposed method is significantly different from the conventional method. However, the discussion regarding the circumstances under which the proposed method can obtain better results when the total calculation amount is limited is not yet fully explored due to the small-scale nature of the experiments. Our future studies will focus on improving and fully discussing this idea.

Traditional eye-tracking systems can be costly and may pose a barrier to entry for researchers interested in studying gaze behavior. In recent years, there have been significant developments in simulating eye-tracking using webcams. However, little research has explored the use of webcam-based eye-tracking data for traditional eye-tracking analysis. In this paper, we propose a webcam-based eye-tracking system that utilizes an dilated convolutional neural networks to detect point of gaze and calculate a range of analysis indicators, such as duration of first fixation and latency of first fixation. By integrating these indicators, we aim to explore the potential of webcam-based eye-tracking for traditional eye-tracking analysis. This approach could significantly reduce the barrier to entry for researchers in the field of gaze behavior research and open up new avenues for studying gaze behavior.

The booming development of Internet-of-Things (IoT) has deployed many IoT systems globally, and this trend is continuously accelerating. However, as many IoT devices are widely deployed, the system-update maintenance is a huge challenge. Over The Air (OTA) update is one promising mechanism for securely updating the firmware of the remote IoT devices.Message Queue Telemetry Transport (MQTT) is one of the most adopted IoT communication protocols globally. It has also been popularly adopted as the communication protocol for delivering the OTA update messages, in addition to delivering normal IoT messages. This paper focuses on MQTT-based OTA models. Even though there exist several MQTT-based OTA models and schemes, we find that no one can simultaneously satisfying user convenience, efficiency and high security. Some sacrifices the privacy against the MQTT broker to achieve user convenience, and some focuses on the privacy while sacrificing the convenience. This paper sorts out the existent models and proposes a new model that distributes the group keys among the manager and the IoT devices, allows the manager deposit the group-key-encrypting firmware on the broker, and then each device can separately access the encrypted OTA images from the broker. We design the scheme using MQTT 5.0 (the new MQTT standard). The analysis and the evaluation show that the new model achieves better privacy protection and gains efficient communication performance.

In this paper, we propose an error decreasing technique using ArUco Marker, in a robot navigation system using SLAM (Simultaneous Localization and Mapping) in which errors occur due to packet loss and time delay. This technique enables more accurate estimation of the position and orientation between the robot and ArUco Marker. Through camera calibration, we convert 3D input values of a real object into undistorted 2D data and establish corresponding relationships between dimensions. Additionally, we use homogeneous transformation matrices to estimate the current direction and degree of rotation of a robot using the marker. Most of robots can reach their destination area through navigation with trial and errors with some time consumption. Therefore, we introduce ArUco Marker to reduce such errors and designed navigation algorithm to enable relatively precise driving with enough fast time. Finally, we compare the navigation accuracy using SLAM of the conventional scheme with the proposed method of twice modifications of the marker information which can reduce the navigation error around actual destination and resulting in accuracy improvement through the position correction process using ArUco Marker recognition.

This study examines the changes that have been made to KNIS since the 2000s, when Korea entered the group of developed countries. Using data on joint research, this study systematically analyzes the interactions among actors representing academia, industry, and government research institutes, and the innovative performance achieved through these interactions. This study argues that the while interactions through joint research generate innovative performance, such interactions have not been occurring as strongly as would be desired, and this has limited the potential for the growth of innovative performance. While reinforcing the capabilities of individual actors is important, this study emphasizes that to build a more effective and systematic NIS, the government should establish policies designed to strengthen the interaction among actors.