Author

H. T. Chang and C. Y. Lin

Abstract

Background

Numerous higher education institutions worldwide have adopted English-language-medium computer science courses and integrated online problem-solving competitions to bridge gaps in theory and practice (Alhamami Education and Information Technologies, 2021; 26: 6549–6562).

Objectives

This study aimed to investigate the factors influencing the use of online competitions in machine learning courses and their impact on student learning. We also analyse disparities in learning outcomes and instructional language effects (Chinese vs. English).

Methods

Among 123 participants at northern Taiwan university, 74 chose Chinese instruction (CMI), and 49 opted for English instruction (EMI). The course spanned 18 weeks: team formation in week one, data analysis, machine learning, and deep learning from week 2 to 8, draft proposals and oral presentations by week 9, instructor guidance in weeks 9–17, followed by off-campus competitions. In week 18, students presented projects for evaluation by judges.

Results

The results showed improved scores in competition proposal writing and oral presentations, especially for CMI students, who excelled in these areas and in terms of creativity. CMI students emphasized domain knowledge, implementation completeness, and technical depth in proposals. The EMI students focused on implementation completeness and artificial intelligence model accuracy, along with creativity.

Conclusion

CMI students achieved superior outcomes in machine learning courses, particularly in terms of competition proposals, oral presentations, and increased creativity. Instructional language choice significantly influenced learning trajectories, leading to distinct knowledge development focuses for CMI and EMI.

@article{chang2024improving, title={Improving student learning performance in machine learning curricula: A comparative study of online problem-solving competitions in Chinese and English-medium instruction settings}, author={Chang, Hui-Tzu and Lin, Chia-Yu}, journal={Journal of Computer Assisted Learning}, year={2024}, publisher={Wiley Online Library} }

Author

H. T. Chang and C. Y. Lin

Abstract

Contribution: This study incorporates competition-based learning (CBL) into machine learning courses. By engaging students in innovative problem-solving challenges within information competitions, revealing that students’ participation in online problem-solving competitions can improve their information technology, and showcase competitions can enhance their competition ability. Background: The CBL model seamlessly integrates project-based learning and competition, placing a strong emphasis on both collective learning and outcomes. This approach cultivates motivation among team members, driving them to enhance their learning and translate knowledge into practical experience. Research Questions: The objective is to examine the disparities in the development of theoretical knowledge, information technology, AI practical ability, and competition ability among students participating in online problem-solving competitions and showcase competitions, and discusses the potential moderating effect of competition type on the relationships between variables in the hypothetical model. Methodology: The study involved 74 students enrolled in machine learning course at a university. The students were given theoretical knowledge and information technology pretests and posttests in the 2nd and 17th weeks, respectively. In the 18th week, the students presented their projects using slideshows and were graded by judges while also submitting their final competition proposal and slides. Findings: Students in online problem-solving competitions can enhance their information technology, while those participating in showcase competitions can improve their competitive ability. Moreover, the competition type was found to moderate the relationships among theoretical knowledge, information technology, and AI model accuracy. The findings suggest that incorporating CBL into machine learning courses effectively cultivates students’ AI practical and competitive abilities.

@article{chang2024applying, title={Applying Competition-Based Learning to Stimulate Students’ Practical and Competitive AI Ability in a Machine Learning Curriculum}, author={Chang, Hui-Tzu and Lin, Chia-Yu}, journal={IEEE Transactions on Education}, year={2024}, publisher={IEEE} }

Author

F. X. R. Chen, C. Y. Lin, H. Y. Siao, C. Y. Jian, Y. C. Yang and, C. L. Lin

Abstract

Defects to popular two-dimensional (2D) transition metal dichalcogenides (TMDs) seriously lower the efficiency of field-effect transistor (FET) and depress the development of 2D materials. These atomic defects are mainly identified and researched by scanning tunneling microscope (STM) because it can provide precise measurement without harming the samples. The long analysis time of STM for locating defects in images has been solved by combining feature detection with convolutional neural networks (CNN). However, the low signal-noise ratio, insufficient data, and a large amount of TMDs members make the automatic defect detection system hard to be applied. In this study, we propose a deep learning-based atomic defect detection framework (DL-ADD) to efficiently detect atomic defects in molybdenum disulfide (MoS2) and generalize the model for defect detection in other TMD materials. We design DL-ADD with data augmentation, color preprocessing, noise filtering, and a detection model to improve detection quality. The DL-ADD provides precise detection in MoS2 (F2-scores is 0.86 on average) and good generality to WS2 (F2-scores is 0.89 on average).

@article{chen2023deep, title={Deep learning based atomic defect detection framework for two-dimensional materials}, author={Chen, Fu-Xiang Rikudo and Lin, Chia-Yu and Siao, Hui-Ying and Jian, Cheng-Yuan and Yang, Yong-Cheng and Lin, Chun-Liang}, journal={Scientific data}, volume={10}, number={1}, pages={91}, year={2023}, publisher={Nature Publishing Group UK London} }

Author

H. T. Chang, C. Y. Lin, W. B. Jheng, S. H. Chen, H. H. Wu, F. C. Tseng, L. C. Wang

Abstract

The objective of this research is based on human-centered AI in education to develop a personalized hybrid course recommendation system (PHCRS) to assist students with course selection decisions from different departments. The system integrates three recommendation methods, item-based, user-based and content-based filtering, and then optimizes the weights of the parameters by using a genetic algorithm to enhance the prediction accuracy. First, we collect the course syllabi and tag each course from twelve departments for the academic years of 2015 to 2020. Next, we use the course tags, student course selection records and grades to train the recommendation model. To evaluate the prediction accuracy, we conduct an experiment on 1490 different courses selected by 5662 students from the twelve departments and then use the root-mean-squared error and the normalized discounted cumulative gain. The results show that the influence of item-based filtering on the course recommendation results is higher than that of user- and content-based filtering, and the genetic algorithm can find the optimal solution and the corresponding parameter settings. We also invite 61 undergraduate students to test our system, complete a questionnaire and provide their grades. Overall, 83.60% of students are more interested in courses at the top of the recommendation lists. The students are more autonomously motivated rather than holding extrinsic informational motivation across the hybrid recommendation method. Finally, we conclude that PHCRS can be applied to all students by tuning the optimal weights for each course selection factor for each department, providing the best course combinations for students’ reference.

@article{chang2023ai, title={AI, Please Help Me Choose a Course}, author={Chang, Hui-Tzu and Lin, Chia-Yu and Jheng, Wei-Bin and Chen, Shih-Hsu and Wu, Hsien-Hua and Tseng, Fang-Ching and Wang, Li-Chun}, journal={Educational Technology \& Society}, volume={26}, number={1}, pages={203--217}, year={2023}, publisher={JSTOR} }

Author

C. Y. Lin, L. C. Wang, and S.P. Chang

Abstract

As the demand of high-speed stream processing grows, in-memory databases are widely used to analyze streaming data. It is challenging for in-memory systems to meet the requirements of high throughput and data persistence at the same time since data are not stored in disks. ARIES logging and command logging are two popular logging methods. In current applications, both ARIES logging and command logging are necessary. However, no checkpointing mechanism includes both the functions of ARIES logging method and command logging method. Besides, adopting ARIES logging method in an in-memory database creates high overhead. Command logging records redundant commands and has high storage cost. To address the above issues, we utilize order-irrelevant characteristics of data structure and incremental checkpointing concepts to devise a data structure based incremental checkpointing (DSIC) mechanism. DSIC mechanism is a very low overhead checkpointing approach while retaining the features of ARIES logging and command logging. DSIC mechanism reduces more than 70 percent logging time of the existing logging scheme and saves 40 percent storage costs of the existing logging scheme.

@article{lin2020incremental, title={Incremental checkpointing for fault-tolerant stream processing systems: A data structure approach}, author={Lin, Chia-Yu and Wang, Li-Chun and Chang, Shu-Ping}, journal={IEEE Transactions on Emerging Topics in Computing}, volume={10}, number={1}, pages={124--136}, year={2020}, publisher={IEEE} }

Author

H.T. Chang, C. Y. Lin, L. C. Wang, and F. C. Tseng

Abstract

In this study, we built a personalized hybrid course recommendation system (PHCRS) that considers students’interests, abilities and career development. To meet students’ individual needs, we adopted the five most widely used algorithms, including content-based filtering, popularity-based methods, item-based collaborative filtering, user-based collaborative filtering, and score-based methods, to build a PHCRS. First, we collected course syllabi and labeled each course (e.g., knowledge/skills taught, basic/advanced level). Next, we used course labels and students’ past course selections and grades to train five recommendation models. To evaluate the accuracy of the system, we performed experiments with students in the Department of Electrical and Computer Engineering, which provides 1794 courses for 925 students and utilizes the receiver operating characteristic curve (ROC) and normalized discounted cumulative gain (NDCG) as metrics. The results showed that our proposed system can achieve accuracies of 80% for ROC and 90% for NDCG. We invited 46 participants to test our system and complete a questionnaire. Overall, 60 to 70% of participants were interested in the recommended courses, while the course recommendation lists produced by content-based filtering were in line with 67.40% of students’ actual course preferences. This study also found that students were more interested in courses at the top of the recommendation lists, and more students were autonomously motivated than held extrinsic informational motivation across the five recommendation methods. These findings highlighted that the proposed course recommendation system can help students choose the courses that interest them most.

@article{chang2022students, title={How Students can Effectively Choose the Right Courses}, author={Chang, Hui-Tzu and Lin, Chia-Yu and Wang, Li-Chun and Tseng, Fang-Ching}, journal={Educational Technology \& Society}, volume={25}, number={1}, pages={61--74}, year={2022}, publisher={JSTOR} }

Author

C. Y. Lin, Y.F. Chiu, L. C. Wang, and D. Niyato

Abstract

Latent tree model (LTM) is a probabilistic tree-structured graphical model, which can reveal the hidden hierarchical causal relations among data contents and play a key role in explainable artificial intelligence. However, because current LTM modeling techniques are only suitable for single-content variable, the applications of LTMs are somewhat limited. Toward this end, a multilayer LTM (ML-LTM) is first presented to deal with the hierarchical clustering issues of multicontent variables. Second, we further develop an ML-LTM-based multicontent recommendation system. Our experiment results show that the proposed ML-LTM can achieve 90% recommendation accuracy, but the current LTM can only has 20%. Third, we propose an incremental update approach for ML-LTM that can save five-sixth updating time comparing with the whole-model retraining approach for achieving the same recommendation accuracy.

@article{lin2020modeling, title={Modeling of multilayer multicontent latent tree and its applications}, author={Lin, Chia-Yu and Chiu, Yu-Fang and Wang, Li-Chun and Niyato, Dusit}, journal={IEEE Transactions on Computational Social Systems}, volume={8}, number={1}, pages={5--19}, year={2020}, publisher={IEEE} }

Author

H.Y. Kung, R.W. Yu, C.H. Chen*, C.W. Tsai, and C. Y. Lin

Abstract

The most important parts of Taiwan's agriculture are animal husbandry and the pig industry, of which the output value reached NT$75.6 billion in 2017. Taiwan has a high technical level of pig raising. However, practical pig-raising skills rely mainly on the inheritance of mentoring experience. The livestock and pig breeding industry in Taiwan has no relevant pig breeding knowledge management information system and no intelligent knowledge question-answering system. Therefore, this study designs and implements an intelligent knowledge question-and-answer system for pig farming. To identify intelligent questions and answers for raising livestock pigs, this study addresses the following issues: (a) to determine the semantic meaning of a sentence, the system needs to accurately interpret the meaning of a question and to identify the expression of a knowledge entity. Therefore, this study applies the lattice long-short-term memory (LSTM) and structured perceptron methods to parse sentences accurately and correctly perform word segmentation. A set of stop words for pig raising in Taiwan is initially established to ensure accurate sentence parsing; (b) to understand the intent of each sentence, the bidirectional gated recurrent init (bi-GRU) method is adopted to realize the knowledge extraction of the livestock in the question and complete the intent detection and slot filling. The bi-GRU extracts the correct livestock knowledge and classifies it into suitable topics; (c) the wide range of knowledge sources in questions often leads to unrelated vocabulary, repeated words, and structural loss in potential answers. To establish an effective knowledge search graph, the method infers implicit knowledge from the existing knowledge and conducts related knowledge retrieval based on the inference results. The knowledge data model is defined in the open standard resource description framework Knowledge filtering and knowledge reasoning strategies are presented to produce candidate answers to the livestock problem; (d) to obtain the final answer, most related works search established knowledge graphs, often producing multiple ambiguous candidate answers. The Siamese neural network (SNN) is adopted to obtain accurate answers. An SNN compares candidate answers to a livestock problem as two input values, previously trained by the bidirectional LSTM neural network. The final answer is determined from the cosine similarity value, which represents the highest relevance to the question. Finally, this study implements the intelligent pig-raising knowledge question-answering system based on the proposed methodology. The evaluation results reveal that the proposed system is accurate and practicable.

@article{kung2021intelligent, title={Intelligent pig-raising knowledge question-answering system based on neural network schemes}, author={Kung, Hsu-Yang and Yu, Ren-Wu and Chen, Chi-Hua and Tsai, Chan-Wei and Lin, Chia-Yu}, journal={Agronomy Journal}, volume={113}, number={2}, pages={906--922}, year={2021}, publisher={Wiley Online Library} }

Author

C. Y. Lin, C. P. Weng, L. C. Wang, H. H. Shuai and W. P. Tseng

Abstract

With the rapid advances in machine learning algorithms and sensing technologies, machine prognostics and health management (PHM) via data-driven approaches has become a trend in sophisticated machine tool industry. The run-to-failure data are necessary for data-driven approaches. However, the average life of the machine is two to three years, the time of collecting data is extended. It is a big challenge to collect run-to-failure data and build a PHM model. Therefore, we propose an Edge-based RNN Anomaly Detection Platform (ERADP). ERADP builds the model based on healthy data and notify anomalies two hours in advance. The true alarm rate is up to 100%. Besides, ERADP can accelerate the training time almost 120 times faster than the traditional model.

@article{lin2019edge, title={Edge-based RNN anomaly detection platform in machine tools}, author={Lin, Chia-Yu and Weng, Chih-Ping and Wang, Li-Chun and Shuai, Hong-Han and Tseng, Wen-Peng}, journal={Smart Science}, volume={7}, number={2}, pages={139--146}, year={2019}, publisher={Taylor \& Francis} }

Author

C. Y. Lin, L. C. Wang and K. H. Tsai

Abstract

In this paper, we present a hybrid real-time incremental stochastic gradient descent (RI-SGD) updating technique for implicit feedback matrix factorization (MF) recommendation systems. Compared with explicit feedback evaluation scores, implicit feedback data are easier to obtain but pose challenges to MF recommendation systems because of the transformation procedures from raw data to user preference scores. Another challenge for MF recommendation systems is the accuracy issue when the speed of the new input data increases. The proposed RI-SGD is designed for computationally-efficient and accurate time-variant implicit feedback MF recommendation system, which consists of alternating least squares with weight regularization in the training phase and stochastic gradient descent in the updating phase. To demonstrate the advantages of the RI-SGD updating technique in terms of computational efficiency and accuracy, we implement the proposed updating techniques in a real-time music recommendation system. Compared with the method of retraining the entire model, our numerical results show that RI-SGD approach can achieve almost the same recommendation accuracy, but requires only about 0.02% of the retraining time.

@article{lin2018hybrid, title={Hybrid real-time matrix factorization for implicit feedback recommendation systems}, author={Lin, Chia-Yu and Wang, Li-Chun and Tsai, Kun-Hung}, journal={Ieee Access}, volume={6}, pages={21369--21380}, year={2018}, publisher={IEEE} }

Author

C. Y. Lin and L. C. Wang

Abstract

Mobile devices become popular with the help of hardware improvements and new functions supported by many sensors. In this paper, we propose a mobile and multi-sensing fusion platform to integrate the unstructured streaming sensing data collecting as well as processing technology and build a QoS performance model to estimate the computing resource of the platform. We also demonstrate three mobile and multi-sensing fusion applications as the examples on the platform. Besides, we discuss the trend and challenges of combining the mobile and multi-sensing fusion technology and signal and information processing in mobile cloud computing in great detail.

@inproceedings{lin2015signal, title={Signal and information processing in mobile cloud computing: Trends and challenges}, author={Lin, Chia-Yu and Wang, Li-Chun}, booktitle={2015 International Conference on Computing, Networking and Communications (ICNC)}, pages={625--629}, year={2015}, organization={IEEE} }

Author

C. Y. Lin, K. H. Tsai, Y. M. Pai, Charles H.-P. Wen, and L.C. Wang

Abstract

Meta-heuristics typically takes long time to search optimality from huge amounts of data samples for applications like communication, medicine, and civil engineering. Therefore, parallelizing meta-heuristics to massively reduce runtime is one hot topic in related research. In this paper, we propose a MapReduce modified cuckoo search (MRMCS), an efficient modified cuckoo search (MCS) implementation on a MapReduce architecture—Hadoop. MapReduce particle swarm optimization (MRPSO) from a previous work is also implemented for comparison. Four evaluation functions and two engineering design problems are used to conduct experiments. As a result, MRMCS shows better convergence in obtaining optimality than MRPSO with two to four times speed-up.

@article{lin2013parallelizing, title={Parallelizing modified cuckoo search on mapreduce architecture}, author={Lin, Chia-Yu and Pai, Yuan-Ming and Tsai, Kun-Hung and Wen, Charles H-P and Wang, Li-Chun}, journal={Journal of Electronic Science and Technology}, volume={11}, number={2}, pages={115--123}, year={2013} }

Author

C. Y. Lin and Y. Z. Chen

Abstract

Deep learning for defect detection has become a critical imperative in contemporary electronics manufacturing. We propose an inpainting-based anomaly detection system to identify defects without labeled defects. An image inpainting model, which discerns disparities between the original and restored versions of the defective image, is designed as the core of our methodology. To further address issues related to reconstructing asymmetric images with defects, we incorporate self-supervised learning (SSL) to extract a broader spectrum of features. In experiments, we compare the proposed method to state-of-the-art models based on MVTec open dataset. Our proposed method can achieve a best performance of 97%, and surpasses the SOTA model by a margin of 57%.

@inproceedings{lin2024inpainting, title={Inpainting-bBased Anomaly Detection System with Self-Supervised Learning}, author={Lin, Chia-Yu and Chen, Yi-Zhen}, booktitle={2024 IEEE International Conference on Industry 4.0, Artificial Intelligence, and Communications Technology (IAICT)}, pages={124--129}, year={2024}, organization={IEEE} }

Author

C. Y. Lin, M. Lu, and Y. C. Yen

Abstract

Artificial Intelligence (AI) has been widely used in manufacturing to detect defects. AI models utilize product images to distinguish whether a product is normal or abnormal. If attackers use the model inversion attack to attack AI models, the input images can be roughly restored, resulting in product information leakage. In this paper, we propose a Privacy-preserving Industrial Defect Detection System (PIDS), which includes three image obfuscation methods to hide input image information and uses them to train the model. Federated learning and differential privacy are also applied to ensure that sensitive data remains decentralized and secure, even during training. Federated learning allows the model to be trained across multiple local datasets without centralized data collection, thereby reducing the risk of data exposure. Differential privacy adds another layer of protection by adding randomness to the learning process, making it hard for attackers to extract sensitive information from the trained model. Experiments show that the proposed system can achieve a high accuracy level of 96.5% in defect image classification. Therefore, the proposed system can detect defects accurately and preserve product information in terms of data and models.

@inproceedings{lin2024privacy, title={Privacy-preserving Federated Learning for Industrial Defect Detection Systems via Differential Privacy and Image Obfuscation}, author={Lin, Chia-Yu and Yeh, Yu-Chen and Lu, Makena}, booktitle={2024 IEEE Conference on Artificial Intelligence (CAI)}, pages={1136--1141}, year={2024}, organization={IEEE} }

Author

P. H. Hung, B. S. Wang, Y. W. Chang, Ted. T. Kuo, L. J. Wang, and C. Y. Lin

Abstract

While deep learning technologies are rising, information security is becoming increasingly important in the modern era. Federated Learning (FL) is a method that allows different clients to collaborate on training deep learning models while ensuring data privacy. However, traditional FL faces the challenge of a single point of failure with the aggregation server if the centralized server fails. Additionally, FL often encounters malicious model attacks, leading to a reduction in global model accuracy. To address these issues, we propose Scale BlockChain Federated Learning (ScaleBCFL) using Hyperledger Fabric (HLF). By leveraging blockchain technology, we decentralize the central server in traditional FL and create multiple shards to reduce communication costs significantly. Furthermore, using a blockchain handler obviates the necessity for users to invest time in learning how to write smart contracts. Instead, users can focus solely on customizing programs related to training, aggregation, and validating models.

@inproceedings{hung2024implementation, title={An Implementation of Blockchain-based Federated Learning Framework}, author={Hung, Po-Hsuan and Wang, Bing-Siang and Chang, Yu-Wei and Kuo, Ted T and Wang, Li-Jen and Lin, Chia-Yu}, booktitle={2024 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={165--166}, year={2024}, organization={IEEE} }

Author

Y. S. Chen, H. Y. Lien, J. Y. Li, and C. Y. Lin

Abstract

Microservice architecture enhances system flexibility and reliability but raises security concerns due to potential malicious attacks. We propose a supervised Out-of-Distribution (OOD) detector leveraging AI and ML to analyze container command sequences. Our technique identifies known and unknown attack patterns, employing out-of-distribution detection. Using a deep neural network, we learn features and minimize classification errors. Comparative evaluations demonstrate its efficacy, aiming to enhance container security and deepen insights into microservice attack behaviors.

@inproceedings{chen2024supervised, title={Supervised Intrusion Detection with Out-of-Distribution Detection for Microservices}, author={Chen, Yong-Syuan and Lien, Hsiang-Yin and Li, Jo-Yu and Lin, Chia-Yu}, booktitle={2024 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={121--122}, year={2024}, organization={IEEE} }

Author

B. Y. Tong, T. T. Kuo, and C. Y. Lin

Abstract

Real-time monitoring of individual node statuses is crucial in distributed systems to promptly identify and address anomalies and enable effective system management. Due to the less intuitive nature of traditional form-based methods for adding charts, there is a growing trend towards integrating large language models (LLMs) into visualization-oriented natural language interfaces (V-NLIs) to automate the entire front-end monitoring process. Although LLMs have advanced significantly, they still exhibit weaker capabilities in code generation, particularly when visualization platforms need LLMs to generate some specialized JSON formats for updating dashboards. To improve the code generation capability of general-purpose LLMs, we propose a system that allows users to create charts using arbitrary natural language. Additionally, concerning the limitations of LLMs in code generation, we aim to leverage the LangChain agent to develop a custom toolkit. The LangChain agent will enable language models to focus on text comprehension while our tool will handle code generation.

@inproceedings{tong2024visualization, title={Visualization-oriented Natural Language Interfaces for Grafana Dashboard}, author={Tong, Bo-Yao and Kuo, Ted T and Lin, Chia-Yu}, booktitle={2024 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={465--466}, year={2024}, organization={IEEE} }

Author

H. Teng, and C. Y. Lin

Abstract

Deepfake has emerged as a significant concern due to its ability to generate fake images and synthesize realistic videos. The increasing development of new techniques for deep-fake creation raises concerns about the cross-forgery issue. Cross-forgery indicates that a model is initially trained to recognize a particular fake and must work against a different unknown forgery. Training a model needs substantial quantities of data, a challenge compounded if the deepfake generation technique is relatively new. Addressing cross-forgery is a critical and essential challenge that requires resolution. In order to solve cross-forgery, our effort presents a method that combines dynamic and static features to identify forgery. For the static component, we extract features similar to general deepfake detection techniques using a single RGB frame as input. Simultaneously, we utilize optical flow analysis to capture changes between consecutive frames for the dynamic part. Our experiments reveal a clear advantage in utilizing combined features, which is particularly evident in cross-forgery scenarios. Specifically, when encountering certain categories, the performance improvement is significant, demonstrating four times better than single-feature models.

@inproceedings{teng2024dynamic, title={Dynamic and Static Features Extraction for Deepfake Detection}, author={Teng, Hao and Lin, Chia-Yu}, booktitle={2024 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={123--124}, year={2024}, organization={IEEE} }

Author

Z. W. Chen, and C. Y. Lin

Abstract

The challenge in defect detection lies in accurately classifying various defect types, each minute in size, and similar defects. Moreover, as defect images are scarce, defect detection becomes a few-shot sample task falling under fine-grained visual classification (FGVC). However, in practical scenarios, discerning defects for model differentiation is complicated by the need to capture subtle features and nuances among different defect classes. Consequently, effective defect detection demands models with robust feature extraction capabilities for few-shot sample tasks to resolve the challenges of FGVC. To address this, we integrate the Self-supervised Knowledge Distillation (SKD) with the few-shot learning model to enhance the feature extraction capabilities of models in handling few-shot FGVC. According to our experiments, our proposed method significantly enhances average accuracy by 5.5%. Hence, this approach enables the rapid training of highly accurate models using limited samples.

@inproceedings{chen2024few, title={Few Shot Ultra Fine-Grained Visual Classification for Defect Detection}, author={Chen, Zheng-Wei and Lin, Chia-Yu}, booktitle={2024 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={119--120}, year={2024}, organization={IEEE} }

Author

P. Y. Li, T. T. Chang, Y. C. Kuo, C. Y. Lin, and H. Y. Chang

Abstract

In the existing anti-money-laundering process, judging abnormal transactions still requires human resources, which is time-consuming and requires companies to pay many human costs. Many experts and scholars have used AI to identify abnormal trading behavior of accounts, but the problem of highly unbalanced data leads to poor model performances. In addition, the complex neural network of deep learning models is considered a black box, which is less likely to explain the model's results. Therefore, our research proposed an "XAI-based AI Anti-Money Laundering Model." We utilize the DNN model to detect laundering, with a recall of 0.94. By applying SHAP to the model, we evaluate the effectiveness of the dataset's ten features on the model. We find that "Payment Format" is the most crucial feature of the anti-money laundering model.

@inproceedings{li2024unveiling, title={Unveiling the Black Box: An XAI-based Anti-Money Laundering Model}, author={Li, Pei-Yi and Chang, Ting-Ting and Kuo, Yu-Chiao and Lin, Chia-Yu and Chang, Heng-Yu}, booktitle={2024 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={293--294}, year={2024}, organization={IEEE} }

Author

Abel C. H. Chen, and C. Y. Lin

Abstract

This study designs a security credential management system, including a certificate authority (CA), a registration authority (RA), and other individuals outside the CA and the RA. An anonymous voting method using elliptic curve cryptography (ECC)-based key expansion has been proposed to achieve anonymity to CA, RA, and individuals. In experiments, the performance of the proposed method based on different security strengths has been evaluated.

@inproceedings{chen2024key, title={Key Expansion Based on Elliptic Curve Cryptography for Anonymous Voting}, author={Chen, Abel CH and Lin, Chia-Yu}, booktitle={2024 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={161--162}, year={2024}, organization={IEEE} }

Author

C. Y. Lin, C. H. Han, W. C. Yin, and Ted. T. Kuo

Abstract

Traditional machine learning typically requires training datasets on local machines or data centers. However, this approach may raise concerns related to data privacy and security. To address these issues, federated learning was proposed. However, federated learning, which involves a server communicating with multiple client devices, can significantly burden the server. Even when using hierarchical federated learning, there is still a considerable cost associated with communication at intermediate nodes. To further alleviate the communication cost burden on intermediate nodes, the most direct approach is to have each intermediate node select a subset of clients for training and accept their model parameters. However, client training data distributions are not uniform, leading to a state known as Non-Independent and Identically Distributed (Non-lID). Unthinkingly selecting clients for training may result in more imbalanced data selection and bias the model training in specific directions. Therefore, we propose the “post-clustering selection”, where clients with similar data distributions are grouped together, and a certain proportion of clients are selected as representatives for training. This approach allows intermediate nodes to reduce communication costs while avoiding the selection of clients with highly imbalanced data distributions. Finally, we integrate differential privacy and secure aggregation to enhance privacy protection and present a framework called 'Cluster-based Privacy-Enhanced Hierarchical Federated Learning Framework with Secure Aggregation (CPE-HFL). From experiments, we reduce the communication volume by up to 29% while maintaining accuracy. Additionally, the accuracy improves more in cases with clustering than those without clustering. The proposed framework can reduce communication costs and effectively protect clients' privacy while maintaining model accuracy.

@inproceedings{lin2024cluster, title={A Cluster-based Privacy-Enhanced Hierarchical Federated Learning Framework with Secure Aggregation}, author={Lin, Chia-Yu and Han, Chih-Hung and Yin, Wei-Chih and Kuo, Ted T}, booktitle={2024 International Conference on Computing, Networking and Communications (ICNC)}, pages={994--999}, year={2024}, organization={IEEE} }

Author

C. Y. Lin, H. Y. Chen, T. M. Chang, and T. J. Wei

Abstract

Mura refers to surface defects or areas of uneven brightness that can occur during factory panel production. Mura can vary in size and shape and be categorized as “light Mura” or “serious Mura.” To optimize the repair process, factories aim to differentiate between the two types of Mura before sending the panels for repair. However, current Mura detection models focus only on identifying “nrmal” and “Mura,” resulting in poor performance in distinguishing between light and serious Mura. To address this issue, we propose an ensemble approach called the Ensemble Image Inpainting and Supervised Modeling Mura Detection System (EISMDS), which combines supervised and image inpainting models to differentiate between the two types of Mura. Experimental results show that our approach improves the True Positive Rate (TPR) by 11 % under a high True Negative Rate (TNR) compared to a single supervised detection model.

@inproceedings{lin2023ensemble, title={An Ensemble of Supervised Learning and Image Inpainting for Mura Detection}, author={Lin, Chia-Yu and Chang, Tzu-Min and Chen, Hao-Yuan and Wei, Tzer-Jen}, booktitle={2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW)}, pages={502--505}, year={2023}, organization={IEEE} }

Author

C. Y. Lin, Y. H. Chou, and Y. C. Cheng,

Abstract

Artificial intelligence (AI) is applied in automated optical inspection (AOI) to help inspect defects and reduce the false discovery rate of AOI in manufacturing industries. In current studies, the training data of AI models are sufficient, and the source data are from the specific production line. However, defect samples are insufficient, and the data source is variant. The current models need more generalization to all machines and take a long training time to build a new model for other appliances. This paper proposes a Deep Learning-based General Defect Detection Framework (DL-GDD) solves the insufficient data issue and the generalization issue of models. We implement a color preprocessing module, a data augmentation module, a data generation module, and four classification models to detect defects and generalize the utilization of DLG-DD. In experiments, we evaluate DLG-DD based on the NEU-CLS and AIdea datasets. The accuracy of DLG-DD is 90%, and the false omission rate and false discovery rate are less than 1%. DLG-DD is a general framework that tackles insufficient data and decreases the false discovery rate of AOI.

@inproceedings{lin2023deep, title={A Deep Learning-based General Defect Detection Framework for Automated Optical Inspection}, author={Lin, Chia-Yu and Chou, Yan-Hung and Cheng, Yun-Chiao}, booktitle={2023 IEEE International Conference on Industry 4.0, Artificial Intelligence, and Communications Technology (IAICT)}, pages={332--337}, year={2023}, organization={IEEE} }

Author

C. Y. Lin, C. L. Li, Y. C. Kuo, Y. C. Cheng, C. Y. Jian, H. T. Huang, and Mitchel M. Hsu

Abstract

Microsectioning is a destructive testing procedure used in the printed circuit board (PCB) fabrication industry to evaluate the quality of PCBs. During cross-section analysis, operators measure PCB component widths manually, which can lead to inconsistencies and make it challenging to establish standardized procedures. We propose a Deep Learning-based Microsection Measurement (DL-MM) Framework for PCB microsection samples to address this issue. The framework comprises four modules: the target detection module, the image preprocessing module, the labeling model, and the coordinate adaptation module. The target detection module is responsible for extracting the area of interest to be measured, which reduces the influence of surrounding noise and improves measurement accuracy. In the image preprocessing module, the target area image is normalized, labeled with coordinates, and resized to different sizes based on the class. The labeling model utilizes a convolutional neural network (CNN) model trained separately for each class to predict its punctuation, as the number of coordinates varies for each class. The final module is the coordinate adaptation module, which utilizes the predicted coordinates to draw a straight line on the expected image for improved readability. In addition, we evaluate the proposed framework on two types of microsections, and the experimental results show that the measurements’ root-mean-square error (RMSE) is only 2.1 pixels. Our proposed framework offers a more efficient, faster, and cost-effective alternative to the traditional manual measurement method.

@inproceedings{lin2023deep, title={A deep learning-based microsection measurement framework for print circuit boards}, author={Lin, Chia-Yu and Li, Chieh-Ling and Kuo, Yu-Chiao and Cheng, Yun-Chieh and Jian, Cheng-Yuan and Huang, Hsiang-Ting and Hsu, Mitchel M}, booktitle={2023 IEEE International Conference on Industry 4.0, Artificial Intelligence, and Communications Technology (IAICT)}, pages={291--294}, year={2023}, organization={IEEE} }

Author

C. Y. Lin, P. F. Lin, W. K. Chung, and Y. H. Lee

Abstract

“Mura” is a phenomenon in which panels have uneven display defects, irregular shapes, and different sizes. It is impossible to produce perfect panels on production lines, so panel inspection is necessary to differentiate between “light Mura” and “serious Mura” manually. The performance of conventional defect detection models for Mura detection is worse since they only differentiate between “normal” and “abnormal” samples. To reduce human cost and increase the accuracy of Mura detection, we propose a “ResUnet-GAN with Dynamic Memory Model an unsupervised anomaly detection method based on a Generative Adversarial Network (GAN) with a memory module to distinguish panel defects. In the dynamic memory, we designed a dynamic feature filtering (DFF) method to choose important features of images, enhancing the ability to recognize light Mura features of the ResUnet-GAN. The proposed model can achieve an Area Under Curve (AUC) of approximately 0.8 for accurate Mura detection. The mechanism of this paper is novel, and the result contributes to practical application.

@inproceedings{lin2023resunet, title={ResUnet-GAN with Dynamic Memory for Mura Defect Detection}, author={Lin, Chia-Yu and Lin, Pin-Fan and Chung, Wei-Kuang and Lee, Yu-Hsien}, booktitle={2023 IEEE International Conference on Industry 4.0, Artificial Intelligence, and Communications Technology (IAICT)}, pages={350--353}, year={2023}, organization={IEEE} }

Author

T. Y. Su, T. T. Kuo, and C. Y. Lin

Abstract

Renewable energy microgrids are flourishing due to the rising environmental consciousness. Blockchain technology is considered a promising solution for advanced metering infrastructures (AMIs) connecting distributed microgrids, enabling power tracking and trading among participants. However, AMIs are vulnerable to attacks, particularly the Inflated Power Generation (IPG) attack, in which energy producers manipulate smart meters to report more energy than they actually produce. Currently, detecting IPG attacks relies mainly on human intervention, which is costly and time-consuming, making the process inefficient a nd n ot s calable. Although b lockchain-based AMIs possess desired features, they are not immune to IPG attacks since there is no on-chain ground truth to assess the reasonability of the reported amount of energy produced. This study proposes a novel Gated Tri-Tower Transformer (GT3) neural network architecture for a multivariate time series classifier t o serve as an IPG attack detector for each AMI blockchain node to validate energy transactions. The GT3 model captures temporal correlations and fuses different features to enhance the accuracy of IPG attack detection. Our experiments with the CAISO dataset demonstrate that the proposed detector outperforms the most advanced detectors, with an increased detection rate from 71.8% to 80.5% and a reduced false alarm rate from 12.2% to 5.6%. Moreover, we investigated and concluded that GT3 performs better than prior works even during the ramp-up periods for as little as one month of collected data.

@inproceedings{syu2023gated, title={Gated Tri-Tower Transformer (GT3)-An Inflated Power Generation Attack Detector for Microgrids}, author={Syu, Ting-Yu and Kuo, Ted T and Lin, Chia-Yu}, booktitle={2023 IEEE International Conference on Blockchain and Cryptocurrency (ICBC)}, pages={1--5}, year={2023}, organization={IEEE} }

Author

C. H. Lin, C. Y. Lin, L. J. Wang, and Ted. T. Kuo

Abstract

We propose a framework for defect detection in production lines that leverages deep learning models, out-of-distribution (OOD) detection, and continual learning to address the challenges of unknown defects and catastrophic forgetting. The proposed method divides classifier training into chronicle tasks, each introducing new defect classes and leveraging OOD detection to classify unknown defects. We evaluate the framework on a highly unbalanced product defect dataset and demonstrated that it outperformed existing approaches, improving the average F-score by 10%. Our method also improve the performance of the PODNet and DER models, but not the WA model due to its poor performance on our dataset. These results suggest that the proposed method has the potential to improve defect detection in production lines, especially for small-quantity-wide-variety production scenarios.

@inproceedings{lin2023continual, title={Continual Learning with Out-of-Distribution Data Detection for Defect Classification}, author={Lin, Cheng-Hsueh and Lin, Chia-Yu and Wang, Li-Jen and Kuo, Ted T}, booktitle={2023 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={337--338}, year={2023}, organization={IEEE} }

Author

Abel C. H. Chen, C. Y. Lin, and Y. T. Lai

Abstract

For improving security, this study proposes a mobile data usage prediction method based on homomorphic encryption (HE)-based similarity calculation without plaintext storage. The cosine similarity, angular similarity, Tanimoto similarity, and Euclidean similarity based on HE according to ElGamal cryptography are proposed and combined in a machine learning method to consider the mobile data usage during the last periods for predicting the mobile data usage during the next period.

@inproceedings{chen2023homomorphic, title={Homomorphic Encryption-Based Similarity Calculation for Mobile Data Usage Prediction}, author={Chen, Abel CH and Lin, Chia-Yu and Lai, Yueh-Ting}, booktitle={2023 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={331--332}, year={2023}, organization={IEEE} }

Author

C. H. Han, W. C. Yin, C. Y. Lin, and Ted. T. Kuo

Abstract

Federated learning is proposed to solve data privacy and security issues for traditional machine learning, which requires the training dataset to be stored locally on a machine or data center for training. However, federated learning may have problems like Non-Independent and Identically Distributed (Non-IID) data and private security. Non-IID can lead to lower training accuracy than expected, and there may be a risk of privacy leakage in the data uploaded by clients. Therefore, this paper proposes CHFDS: Clustered-based Hierarchical Federated Learning Framework with Differential Privacy and Secure Aggregation. Before training begins, we cluster all clients so that the data distribution between clients in each group is similar. This means only a random subset of clients from each cluster is selected in each training round instead of all clients participating in the training. We can use this method to adjust the data balance of participating training. Finally, we add differential privacy and secure aggregation to the clustering and training process to improve the privacy and security of the proposed clustered federated learning framework.

@inproceedings{han2023chfds, title={CHFDS: Clustered-based Hierarchical Federated Learning Framework with Differential Privacy and Secure Aggregation}, author={Han, Chih-Hung and Yin, Wei-Chih and Lin, Chia-Yu and Kuo, Ted T}, booktitle={2023 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={215--216}, year={2023}, organization={IEEE} }

Author

T. M. Chang, H. Y. Chang, and C. Y. Lin

Abstract

Mura is usually caused by inhomogeneity and material defects in the manufacturing process. According to the JND value, it can be divided into light Mura and serious Mura. In order to optimize the repair process, the factory hopes to distinguish between light Mura and serious Mura before sending them to the repair site. However, the traditional AI model only distinguishes between normal and Mura and is ineffective in distinguishing between light Mura and serious Mura. To address this issue, we propose a Mura Detection System using an image inpainting model with a self-supervised technique and an attention module to distinguish light Mura and serious Mura. The experiment results show that the proposed method’s Area Under Curve (AUC) can reach 0.854.

@inproceedings{chang2023image, title={Image Inpainting with Self-Supervised Learning for Mura Detection System}, author={Chang, Tzu-Min and Chen, Hao-Yuan and Lin, Chia-Yu}, booktitle={2023 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={339--340}, year={2023}, organization={IEEE} }

Author

M. C. Luo and C. Y. Lin

Abstract

The advent of intelligent transportation systems has led to the direct recognition of license plates by license plate recognition systems. However, the system is affected by environmental factors, such as underexposure, resulting in over-dark images. In this paper, we propose a two-stage restoration system to address these issues. In the first stage, we use image processing to enhance the license plate, and in the second stage, we employ a deep learning model to further repair the license plate. Our approach has greatly improved the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) metrics, increasing by 3.8094 and 0.0137.

@inproceedings{luo2023license, title={License Plate Restoration System}, author={Luo, Meng-Chen and Lin, Chia-Yu}, booktitle={2023 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={333--334}, year={2023}, organization={IEEE} }

Author

Y. C. Kuo, C. L. Li, W. C. Yin, and C. Y. Lin

Abstract

With the progress of technology, most of the entertainment of the younger generation is related to electronic products. People gradually forget traditional games. Red Light, Green Light has ample liquidity that people worldwide know it. Therefore, to bring traditional games back to children’s childhood, we use IoT devices to develop a Red Light, Green Light Game System. This system comprises three modules: an identification module, a distance measurement module, and a notification module. In the identification module, we use the Raspberry Pi equipped with the MediaPipe Pose to detect 33 pose nodes of the human body. Then the model reads the camera image through OpenCV and marks the players’ skeleton position. Finally, we compare the players’ skeleton positions in the front and rear images to determine whether the player has moved. The ultrasonic sensor "HC-SR04" is used in the distance measurement module to measure the distance. The change of the front and rear distances is used to identify whether a player is passing through the finish line. The notification module places the "TM1637" four-digit display and the speaker. It will display the remaining time and play the sounds needed for the game during the gameplay. The proposed game system allows adults to relive their childhood and introduces traditional games to younger generations.

@inproceedings{kuo2023red, title={Red Light, Green Light Game System based on IoT Devices}, author={Kuo, Yu-Chiao and Li, Chieh-Ling and Yin, Wei-Chih and Lin, Chia-Yu}, booktitle={2023 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={767--768}, year={2023}, organization={IEEE} }

Author

Y. C. Cheng, Y. H. Chou, and C. Y. Lin

Abstract

The blooming season is a crucial aspect of tourism in Taiwan, but it is subject to annual variations caused by weather factors such as rainfall and temperature. While many AI models are on the market for predicting flowering, they often need more applicability to specific regions due to climate variations. Moreover, Taiwan’s climate is known for being changeable, which can further complicate flower prediction. Using Taiwan’s climate and flowering date as training parameters, our model can achieve significantly higher accuracy than models that do not incorporate Taiwan’s climate information. This paper presents an App called "Booming Blooming," which integrates a flower prediction model with real-time weather information. The App utilizes weather data from Taiwan’s Central Weather Bureau to predict the optimal time for flower viewing and provides users with up-to-date weather forecasts. With this App, users can plan their flower-viewing trips more effectively. Moreover, the App includes a built-in Google map to assist users in locating nearby stores, traffic conditions, and other people at popular flower-viewing locations. Additionally, Booming Blooming offers a flower-sharing platform where users can share the latest information on flower blooming conditions. Overall, the proposed flower blooming prediction model and App provide a convenient and efficient way for Taiwanese to enjoy flower-viewing activities.

@inproceedings{cheng2023booming, title={Booming Blooming: When Can You Enjoy Flowering Season?}, author={Cheng, Yun-Chiao and Chou, Yan-Hung and Lin, Chia-Yu}, booktitle={2023 International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan)}, pages={335--336}, year={2023}, organization={IEEE} }

Author

J. J. Gung, C. Y. Lin, P. F. Lin, and W. K. Chung

Abstract

Defect detection is essential in production lines to guarantee the quality of products. However, detecting tiny defects is difficult. Besides, as the variety of products increases, the variety of defects also increases. Models take much time to retrain. In this paper, we propose an “Incremental Meta Defect Detection (IMDD) System,” which utilizes incremental meta-learning to detect tiny defects. We decompose the model into feature pyramids and use feature alignment to improve the sensitivity of minor defects. Incremental learning utilizes knowledge distillation but this affects the learning of new categories, so the model is quickly adapted to new categories. We further combine incremental learning with meta-learning to increase the generality of the model. In experiments, the proposed model is 1.14 times more accurate than previous techniques. Therefore, the proposed system can enhance the ability to identify minor defects and quickly adapt to new defect types.

@inproceedings{gung2022incremental, title={An incremental meta defect detection system for printed circuit boards}, author={Gung, Jia--Jiun and Lin, Chia--Yu and Lin, Pin--Fan and Chung, Wei--Kuang}, booktitle={2022 IEEE International conference on consumer electronics-Taiwan}, pages={307--308}, year={2022}, organization={IEEE} }

Author

M. M. Hsu, M. H. Wu, Y. C. Cheng, and C. Y. Lin

Abstract

The serial number is essential for product tracking in industries. Without serial numbers, the control of production is lost. Optical character recognition (OCR)—a technique for extracting characters' features and utilizing these features to recognize them in digital images—is widely used to identify serial numbers and improve product tracking efficiency. However, it is relatively challenging for OCR to adapt to digital images containing unclean environments in industries. Images taken in such environments can include various objects, making it hard for OCR models to focus on serial numbers. In this paper, we propose an industrial product serial number recognition framework, which can efficiently adapt complex structures of the background environment and recognize serial numbers in industries. The framework includes the data preprocessing stage, detection stage, and recognition stage. Eventually, our framework reaches nearly four times better on the testing dataset than the baseline score and increases the efficiency of product tracing in industries.

@inproceedings{hsu2022efficient, title={An Efficient Industrial Product Serial Number Recognition Framework}, author={Hsu, Mitchel M and Wu, Ming-Hsien and Cheng, Yun-Chieh and Lin, Chia-Yu}, booktitle={2022 IEEE International Conference on Consumer Electronics-Taiwan}, pages={263--264}, year={2022}, organization={IEEE} }

Author

S. Q. Ye, C. S. Xue, C. Y. Jian, Y. Z. Chen, J. J. Gung, and C. Y. Lin

Abstract

Automated optical inspection (AOI) is essential in the electronic manufacturing production line. Strict screening rules lead to a high false alarm rate of AOI. Many industries use AI models to classify defects. The lack of flawed data and the uneven distribution of categories is a big challenge for model training. Furthermore, the AI model must be retrained when adding new production line data, and the time cost is high. In order to reduce the false alarm rate and improve the generalization of the AI model, we build a deep learning- based generic solder defect detection system (GSDD) to classify defects into seven types. In GSDD, the color gradation adjustment module solves the problem of color difference, and the data augmentation module solves the problem of variable data. In the experiment, we use the data set provided by the enterprise to evaluate the accuracy of the model to 96%, and the model can be applied to different machines. Thus, GSDD is a general model and can efficiently detect defects.

@inproceedings{ye2022deep, title={A deep learning-based generic solder defect detection system}, author={Ye, Shi-Qi and Xue, Chen-Sheng and Jian, Cheng-Yuan and Chen, Yi-Zhen and Gung, Jia-Jiun and Lin, Chia-Yu}, booktitle={2022 IEEE International Conference on Consumer Electronics-Taiwan}, pages={99--100}, year={2022}, organization={IEEE} }

Author

H. Y. Chen, Y. C. Yeh, Makena Lu, and C. Y. Lin

Abstract

There have been many related security issues about Artificial Intelligence (AI) in recent years. During the manufacturing process, products are captured by images for defect detection. If attackers use the model inversion attack to attack the AI model, the input image can be roughly restored, resulting in product information leakage. In this paper, we propose a system that confuses input images and uses them to train the model. Experiments show that our model has a high accuracy of 94.4% in defect image classification. Thus, the proposed system can achieve product information protection and accurate defect detection.

@inproceedings{chen2022image, title={Image Confusion Applied to Industrial Defect Detection System}, author={Chen, Hao-Yuan and Yeh, Yu-Chen and Lu, Makena and Lin, Chia-Yu}, booktitle={2022 IEEE International Conference on Consumer Electronics-Taiwan}, pages={463--464}, year={2022}, organization={IEEE} }

Author

H. T. Huang, T. Y. Chiu, and C. Y. Lin

Abstract

Recently, many factories have utilized AI to help Automatic Optical Inspection (AOI) machines accurately detect defects. They usually deploy AI models on the clouds and submit the data to the clouds for inference. However, transmission delay increases the response time of the AI model. If AI can differentiate defects on the local edge devices, the production efficiency can be significantly improved. In this paper, we propose a light-weight defect detection system that utilizes pruning techniques to compress the model and can accurately detect defects at a faster speed. Besides, we compare the performance of pruned and unpruned models on Kneron KL520 AI dongle and NVIDIA Jetson Nano to verify the superior ability of pruning to accelerate inference. The accuracy of the pruned model in the proposed system can reach 97.7% on Kneron KL520 AI dongle. The inference speed is 28.2 frames per second, 1.6 times faster than the unpruned model. Also, compared to NVIDIA Jetson Nano, the inference speed on Kneron KL520 AI dongle is two times faster. This result shows the better performance of Kneron KL520 AI dongle than NVIDIA Jetson Nano on inference. In summary, the proposed system can significantly improve the efficiency of production lines and avoid the information security risks brought by cloud computing.

@inproceedings{huang2022light, title={A Light-Weight Defect Detection System for Edge Computing}, author={Huang, Hsiang-Ting and Chiu, Tzu-Yi and Lin, Chia-Yu}, booktitle={2022 IEEE International Conference on Consumer Electronics-Taiwan}, pages={521--522}, year={2022}, organization={IEEE} }

Author

C. Y. Lin, F. M. Kang, P. M. Hsu, J. J. Chen and Y. C. Tseng

Abstract

With the maturity of 5G networks and V2X technologies, utilizing V2X to increase driving safety becomes feasible. Road side units (RSU) recognize dangerous events and notify the related vehicles using 5G V2X communications. The alert latency between RSUs and related vehicles may cause car accidents. Besides, the V2X emergency alarms are broadcasted by dipole antennas. Many unrelated vehicles will receive unnecessary alerts and distract drivers. This paper proposes an accurate vehicle-to-vehicle instant alert system (V2V-IAS) to detect potential dangerous events and alert target vehicles by directional antennas to reduce transmission delay and interference. In V2V-IAS, a vehicle can fuse camera information and cooperative awareness messages (CAM) messages from surrounding vehicles carried to predict dangerous events without relying on RSUs. Once an event is detected, the vehicle will send an alert to specific cars in real time using directional antennas to increase the transmission success rate and reduce disturbing unrelated vehicles. We also use two sets of Qualcomm MDM9150 to conduct real road tests and simulate multi-vehicle environments on Matlab to validate our idea. From experiments, directional antennas can reduce the possibility of packet collisions up to 21% and transmit at the same distance with less output power compared to dipole antennas.

@inproceedings{lin2021accurate, title={An accurate vehicle-to-vehicle instant alert system using directional antennas}, author={Lin, Chia-Yu and Kang, Fu-Ming and Hsu, Po-Min and Chen, Jen-Jee and Tseng, Yu-Chee}, booktitle={2021 IEEE 94th Vehicular Technology Conference (VTC2021-Fall)}, pages={1--5}, year={2021}, organization={IEEE} }

Author

Y. C. Cheng, H. T. Chang, C. Y. Lin, and H. Y. Chang

Abstract

Peer-to-peer (P2P) lending provides borrowers with relatively low borrowing interest rates and gives lenders a channel for investment on an online platform. Since most P2P lending does not require any guarantees, the overdue payment of borrowers results in a massive loss of lending platforms and lenders. Many risk prediction models are proposed to predict credit risk. However, these works build models with more than 50 features, which causes a lot of computation time. Besides, in most P2P lending datasets, the number of non-default data far exceeds the number of default data. These researches ignore the data imbalance issue, leading to inaccurate predictions. Therefore, this study proposes a credit risk prediction system (CRPS) for P2P lending to solve data imbalance issues and only require few features to build the models. We implement a data preprocessing module, a feature selection module, a data synthesis module, and five risk prediction models in CRPS. In experiments, we evaluate CRPS based on the de-identified personal loan dataset of the LendingClub platform. The accuracy of the CRPS can achieve 99%, the recall reaches 0.95, and the F1-Score is 0.97. CRPS can accurately predict credit risk with less than 10 features and tackle data imbalance issues.

@inproceedings{cheng2021predicting, title={Predicting credit risk in peer-to-peer lending: A machine learning approach with few features}, author={Cheng, Yun-Chieh and Chang, Hui-Ting and Lin, Chia-Yu and Chang, Heng-Yu}, booktitle={2021 International Conference on Technologies and Applications of Artificial Intelligence (TAAI)}, pages={295--300}, year={2021}, organization={IEEE} }

Author

C. Y. Lin, T. T. Chen, L. C. Wang, and H. H. Shuai

Abstract

Neural network (NN)-based anomaly-detection model has achieved unprecedented performance in internet of things such as fault diagnosis for machine tools. To train a promising neural network-based fault diagnosis model, sufficient fault data is necessary. In reality, machines are rarely operating in fault and cause an unbalance between health data and fault data. Various Generative adversarial networks (GANs) such as auxiliary classifier GAN (ACGAN) were proposed to generate artificial data that are similar to real data for data augmentation. However, applying the current GAN in fault diagnosis model, the fake data makes decision boundary unstable and results in low classification accuracy. Besides, a certain amount of fault data is required and only one-to-one domain generation can be achieved, rather than one-to-many. In this paper, we propose an end-to-end fault diagnosis framework with health-based fault generative adversarial network (HFGAN) to solve data imbalance issue, augmentation uncertainty issue and improve fault diagnosis for machine tools. Based on the proposed HFGAN, various high quality target domain fault data is generated based on health data and oneto-many data generation is achieved. From the experiments, mixing the generated fake data with 7.5% real data as the input of the fault diagnosis model, the classification accuracy achieves 99.8%. Moreover, the cost of data collection in factories can be reduced since only 7.5% of real machine data is required for fault diagnosis.

@inproceedings{lin2020health, title={Health-based fault generative adversarial network for fault diagnosis in machine tools}, author={Lin, Chia-Yu and Chen, Tzu-Ting and Wang, Li-Chun and Shuai, Hong-Han}, booktitle={The 4th International Workshop on Artificial Intelligence of Things}, year={2020} }

Author

C. H. Chen, Y. H. Zhang, W. H. Guo, M. Y. Pan, L. J. Lyu, C.Y. Lin

Abstract

This study proposes a ship recognition system which includes intelligent bridge piers and a ship recognition server. The ship recognition server can analyse the contour features of ship images from intelligent bridge piers by the proposed contour accentuation method; the ship image with contour accentuation can be adopted as the inputs of transfer learning-based neural network for ship classification by the proposed transfer learning-based ship recognition method. In practical experiments, the results showed that the proposed transfer learning-based ship recognition method with contour accentuation can obtain higher accuracy, and the accuracy of the proposed method was 97.79%.

@inproceedings{chen2020contour, title={Contour accentuation for transfer learning-based ship recognition method}, author={Chen, Chi-Hua and Zhang, Yizhuo and Guo, Wenzhong and Pan, Mingyang and Lyu, Lingjuan and Lin, Chia-Yu}, booktitle={Companion Proceedings of the Web Conference 2020}, pages={61--62}, year={2020} }

Author

C. Y. Lin, L.H. Ho and H.S. Huang

Abstract

Author

K.H. Tsai, C. Y. Lin and L. C. Wang

Abstract

With the prosperity of e-commerce, on-line vendors use recommendation systems in different fields. Classic recommendation algorithms are designed assuming that data is stationary and will not change over time. However, since the scale and variability of data are growing gradually, these methods will encounter the issues of the memory deficient and the out-of-date model, which degrade the recommendation accuracy intensively. In addition, retraining the whole model for every new arrival record results in high complexity. In this paper we propose a light-weight adaptive updating method to overcome these issues. Comparing with the explicit feedback recommendation, which asks the customers to express their opinions on the recommended items, the implicit feedback recommendation is easier to collect and non- intrusive way. However, the dynamic time-variant system with implicit feedback has not been seen in the literature. In this paper, we propose a real- time incremental updating algorithm (RI-SGD) to deal with time-variant systems based on the implicit feedback. We compare our method with methods that retraining the whole model and show that our method costs less than 1% of the retraining time with a competitive accuracy.

@inproceedings{tsai2015reconstruct, title={Reconstruct dynamic systems from large-scale open data}, author={Tsai, Kun-Hung and Lin, Chia-Yu and Wang, Li-Chun and Chen, Jian-Ren}, booktitle={2015 IEEE Global Communications Conference (GLOBECOM)}, pages={1--6}, year={2015}, organization={IEEE} }

Author

C. Y. Lin and L. C. Wang

Abstract

improvements and new functions supported by many sensors. In this paper, we propose a mobile and multi-sensing fusion platform to integrate the unstructured streaming sensing data collecting as well as processing technology and build a QoS performance model to estimate the computing resource of the platform. We also demonstrate three mobile and multi-sensing fusion applications as the examples on the platform. Besides, we discuss the trend and challenges of combining the mobile and multi-sensing fusion technology and signal and information processing in mobile cloud computing in great detail.

@inproceedings{lin2015signal, title={Signal and information processing in mobile cloud computing: Trends and challenges}, author={Lin, Chia-Yu and Wang, Li-Chun}, booktitle={2015 International Conference on Computing, Networking and Communications (ICNC)}, pages={625--629}, year={2015}, organization={IEEE} }

Author

Y. J. Chen, C. Y. Lin, B. Lin, and L. C. Wang

Abstract

Author

C. Y. Lin, Z.F. Jiang, L. C. Wang and B.S. Paul Lin

Abstract

Advertisements and coupon forwarding among mobile devices and social platforms become significant in recent years. However, when users are in the crowded places where the network connection is unstable such as shopping malls, forwarding of advertisements to target users in limited time is challenge. Thus, we propose an interest-aware opportunistic advertising by mining social and consuming information system (ISC) and algorithm. In ISC system, we combine social relationship and consuming information to find out the consumers who have similar interest. ISC algorithm adopts the result of ISC system to reduce the delay of broadcasting advertisement and increase the advertisement effect. We also implement and evaluate the performance of ISC system and algorithm on mobile devices and the ONE simulator. Our experiment results show that ISC algorithm is cost effective and intensively magnify the advertising effeteness with many customers under uncertain network and time-limited environments.

@inproceedings{lin2014effective, title={An effective algorithm for interest aware opportunistic advertising by mining social and consuming information}, author={Lin, Chia-Yu and Jiang, Zhi-Feng and Wang, Li-Chun and Lin, Bao-Shuh Paul}, booktitle={2014 IEEE 79th Vehicular Technology Conference (VTC Spring)}, pages={1--5}, year={2014}, organization={IEEE} }

Author

Y. J. Chen, C. Y. Lin, and L. C. Wang

Abstract

Communication networks such as cellular phone networks are quite likely to be severely damaged during a large-scale of disaster, making SOS message dissemination to rescue authorities extremely difficult. In this paper, we first propose a FM radio-based emergency communication service through the integration of FM transmitters into smartphones. FM radio provides a number of advantages such as longer propagation length and less susceptible to obstacles, making it suitable for broadcasting SOS messages. Besides, we design Morse code-based SOS message dissemination using FM transmitters (MCSOS-FM) and its corresponding communication procedure for the emergency communication service in order to improve communication range, victim localization and evacuation route planning. The combination of FM radio and Morse code is completely compatible with the current radio system equipped by rescue workers, and thus increases the possibilities of successful receiving and recognizing SOS messages. Our experiment result based on smartphone implementation shows that the proposed emergency communication service is cost effective and energy efficient with relatively large communication range.

@inproceedings{chen2013personal, title={A personal emergency communication service for smartphones using FM transmitters}, author={Chen, Yu-Jia and Lin, Chia-Yu and Wang, Li-Chun}, booktitle={2013 IEEE 24th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC)}, pages={3450--3455}, year={2013}, organization={IEEE} }

Author

Y. J. Chen, C. Y. Lin, and L. C. Wang

Abstract

In this paper, we develop a privacy preserving mechanism for a scheduling service for a group of users, which can compute the joint available time based on users' calendar information. We also propose a weighted voting game with partially homomorphic encryption (WVGPHE) scheme to protect the privacy of users' calendar information as well as users' weight. Even with the privacy protection, our experimental results show that the proposed scheme is still very computational and bandwidth efficient enough to provide immediate group meeting in mobile communications environments.

@inproceedings{chen2013privacy, title={A Privacy-Preserved Joint Group Time Scheduling Mechanism for Mobile Social Applications}, author={Chen, Yu-Jia and Lin, Chia-Yu and Wang, Li-Chun}, booktitle={2013 12th IEEE International Conference on Trust, Security and Privacy in Computing and Communications}, pages={150--155}, year={2013}, organization={IEEE} }

Author

Y. J. Chen, C. Y. Lin, and L. C. Wang

Abstract

In this paper, we propose a sensors-assisted rescue service architecture to integrate rescue schemes for different purposes, including disaster prediction, evacuation planning, and emergency broadcast. In the proposed architecture, multiple-sensed mobile devices are designed to provide a personalized situational awareness, thereby further enhancing the flexibility and efficiency of rescue services. Reliability and scalability of rescue services are improved by leveraging the dynamical resource provision of cloud computing. The proposed rescue service architecture is implemented to show the advantages of power efficiency and scalability of the proposed rescue service architecture.

@inproceedings{chen2013sensors, title={Sensors-assisted rescue service architecture in mobile cloud computing}, author={Chen, Yu-Jia and Lin, Chia-Yu and Wang, Li-Chun}, booktitle={2013 IEEE Wireless Communications and Networking Conference (WCNC)}, pages={4457--4462}, year={2013}, organization={IEEE} }

Author

C. Y. Lin, Y. J. Chen, L. C. Wang and Y. C. Tseng

Abstract

Smart phones with touch screens have become very popular and have changed our behaviors of using handsets. However, using touch screens is not safe for mobile phone users especially when they are driving cars. Thus, many applications using smart phones cannot be initiated because users who are driving the cars cannot easily touch the small icons on the screens of smart phones. To overcome this issue, we propose a proximity sensors based "no-touch" mechanism for smart phones by applying proximity sensors to initiate mobile applications without the need of touching the screen. We will discuss how to implement the proximity-sensors driven "no-touch" mechanism in Android platform and investigate its performance issues regarding detection accuracy and power consumption. Speech-oriented applications using the proposed "no-touch" mechanism on Android is also demonstrated in this paper.

@inproceedings{lin2012proximity, title={A proximity sensor based no-touch mechanism for mobile applications on smart phones}, author={Lin, Chia-Yu and Chen, Yu-Jia and Wang, Li-Chun and Tseng, Yu-Chee}, booktitle={2012 IEEE Vehicular Technology Conference (VTC Fall)}, pages={1--5}, year={2012}, organization={IEEE} }

Author

C. Y. Lin, Y. A. Chen, Y. C. Tseng and L. C. Wang

Abstract

In cloud computing environments, users can rent virtual machines (VMs) from cloud providers to execute their programs or provide network services. While using this kind of cloud services, one of the biggest problems for the users is to determine the proper number of VMs to complete the jobs considering both budget and time. In this paper, we propose a resource prediction framework (RPF), which can help users choose the minimum number of virtual machines to complete their jobs within a user specified time constraint. In order to verify the feasibility of RPF, we have done three case studies, namely parallel frequent pattern growth (FP-Growth), parallel K-means, and Particle Swarm Optimization (PSO). FP-growth, K-means and PSO are data intensive algorithms. These algorithms are typically executed repeatedly with different execution parameters to find the optimal results. When evaluating RPF by these algorithms in cloud environments, we have to modify them to parallel versions. The evaluation results indicate that RPF can successfully obtain the minimum number of VMs with acceptable errors. According to our case studies, the proposed RPF can be adopted by data intensive jobs by providing flexibility to both end users and cloud system providers.

@inproceedings{lin2012flexible, title={A flexible analysis and prediction framework on resource usage in public clouds}, author={Lin, Chia-Yu and Chen, Yan-Ann and Tseng, Yu-Chee and Wang, Li-Chun}, booktitle={4th IEEE International Conference on Cloud Computing Technology and Science Proceedings}, pages={309--316}, year={2012}, organization={IEEE} }

Author

C. Y. Lin, Y. J. Chen, L. C. Wang and Y. C. Tseng

Abstract

Author

C. Y. Lin, Y. J. Chen, L. C. Wang and Y. C. Tseng

Abstract

Author

C. Y. Lin, Y. J. Chen, Y. M. Chang and L. C. Wang

Abstract

Author

Abstract