Search

Recently, the world has been dealing with a severe outbreak of COVID-19. The rapid transmission of the virus causes mild to severe cases of cough, fever, body aches, organ failures, and death. An increasing number of patients, fewer diagnostic options, and extended waiting periods for test results all put pressure on healthcare systems, increasing the virus’s spread. A concise and accurate automatic diagnosis is crucial to identify infected patients in the early stage. This paper proposes a machine learning-based predictive framework to identify COVID-19 cases from clinical data using an optimized union ensemble feature selection (OUEFS) approach. The OUEFS is based on the union ensemble of the feature subsets obtained through a rigorous feature selection (FS) process. It also involves a performance optimization of the ML classifiers. Initially the OUEFS identified key features from the publicly accessible COVID-19 dataset using FS methods such as Mutual Information Feature Selection (MIFS), Recursive Feature Elimination (RFE), and the RidgeCV. The most important features were selected using Top-k thresholding technique. Then selected subsets of features were integrated using a union ensemble approach where an optimal combination of features with enhanced predictive power is derived. This composite feature set was subsequently utilized for model training and evaluation. The classification was conducted using machine learning algorithms such as linear SVM, gradient boosting (GB), logistic regression (LR), and Adaboost to compare their effectiveness on individual and combined feature subsets. We also conducted a Genetic Algorithm (GA) based hyperparameter optimization (HPO) which further refined our training process and enhanced the accuracy of our proposed approach. Experimental results show that the union ensemble of MIFS and RidgeCV FS techniques and the Adaboost classifier and GA HPO achieved 96.30% accuracy. Our optimized union ensemble approach demonstrated superior performance over previous ensemble-based approaches to predict COVID-19 disease, thus offering a robust tool for early and efficient diagnosis without requiring hospital visits.

Federated Learning (FL) has emerged as a promising paradigm for privacy-preserving distributed Machine Learning (ML), enabling model training across distributed devices without compromising data privacy. However, the impact of hyper-parameters on FL model performance remains understudied and most of the existing FL studies rely on default or out-of-the-box hyper-parameters, often leading to suboptimal convergence. This study specifically investigates the intricate relationship between key hyper-parameters—learning rate, epochs per round, batch size, and client participation ratio (CPR)—and the performance of FL models on two distinct datasets: CIFAR-10 using ResNet-18 and FashionMNIST using a simple CNN model. Through systematic exploration on these datasets, employing a centralized server and 200 clients, we elucidate the significant impact of varying hyper-parameters. Our findings underscore the importance of dataset-specific hyper-parameter optimization, revealing contrasting optimal configurations for the complex CIFAR-10 dataset and the simpler FashionMNIST dataset. Additionally, the correlation analysis offers a deep understanding of hyper-parameter inter-dependencies, essential for effective optimization. This study provides valuable insights for practitioners to customize hyper-parameter configurations, ensuring optimal performance for FL models trained on different types of datasets and provides a foundation for future exploration in hyper-parameter optimization within the FL domain.

Cyberbullying has emerged as a pervasive issue in the digital age, necessitating advanced techniques for effective detection and mitigation. This research explores the integration of word embeddings, emotional features, and federated learning to address the challenges of centralized data processing and user privacy concerns prevalent in previous methods. Word embeddings capture semantic relationships and contextual information, enabling a more nuanced understanding of text data, while emotional features derived from text extend the analysis to encompass the affective dimension, enhancing cyberbullying identification. Federated learning, a decentralized learning paradigm, offers a compelling solution to centralizing sensitive user data by enabling collaborative model training across distributed devices, preserving privacy while harnessing collective intelligence. In this study, we conduct an in-depth investigation into the fusion of word embeddings, emotional features, and federated learning, complemented by the utilization of BERT, Convolutional Neural Networks (CNN), Deep Neural Networks (DNN), and Long Short-Term Memory (LSTM) models. Hyperparameters and neural architecture are explored to find optimal configurations, leading to the generation of superior results. These techniques are applied in the context of cyberbullying detection, using publicly available multi-platform (social media) cyberbullying datasets. Through extensive experiments and evaluations, our proposed framework demonstrates superior performance and robustness compared to traditional methods. The results illustrate the enhanced ability to identify and combat cyberbullying incidents effectively, contributing to the creation of safer online environments. Particularly, the BERT model consistently outperforms other deep learning models (CNN, DNN, LSTM) in cyberbullying detection while preserving the privacy of local datasets for each social platform through our improved federated learning setup. We have provided Differential Privacy based security analysis for the proposed method to further strengthen the privacy and robustness of the system. By leveraging word embeddings, emotional features, and federated learning, this research opens new avenues in cyberbullying research, paving the way for proactive intervention and support mechanisms. The comprehensive approach presented herein highlights the substantial strengths and advantages of this integrated methodology, setting a foundation for future advancements in cyberbullying detection and mitigation.

In this paper, we address the challenge of limited resources in Internet of Things (IoT) devices by proposing a solution based on digital twin in distributed edge computing networks. Edge computing is a promising approach that moves computing resources closer to the network’s edge to reduce response times in IoT applications. However, simply offloading tasks from IoT devices to edge computing does not accelerate user control. To enhance task performance and improve user management experience, we introduce optimized task scheduling and virtual object management based on a digital twin concept. Our system incorporates virtualization, synchronization, visualization, and simulation functionalities to provide digital twin capabilities. Additionally, we develop a user-friendly web application with a graphical user interface (GUI) for intuitive management of edge computing services. To support our approach, we implement an edge computing supervisor that generates virtualized objects such as edge gateways, IoT devices, and services. These virtual objects serve as resources for creating tasks. Using our proposed digital twin platform, users can dynamically create new tasks based on demand, easily deploy and execute tasks in specific locations, and dynamically allocate edge network resources according to task requirements. An optimized task scheduling mathematical model is presented to compare task scheduling done with and without optimization. Further, the edge computing and digital twin based optimized task scheduling method is integrated with Federated Learning for collaborative learning and privacy preserved computation of sensors sensitive data. We demonstrate the effectiveness of our system by generating tasks for data collection related to indoor environment for prediction of Predicted Mean Vote (PMV) for thermal comfort index of smart homes occupants using HTTP and IoTivity-based devices in distributed edge computing networks. These tasks are properly delivered and executed on the expected edge gateways, showcasing the successful integration of our digital twin platform with edge computing networks. Further, the optimized task scheduling has improved the overall performance of the proposed system, keeping in view latency and processing time.

Simulation and optimization methods have been widely used in forklift design due to their cost-effectiveness. However, this type of method involves challenges such as the accuracy of the simulation model and the simulation solution time. These challenges reduce the stability and precision of the surrogate model and hence generate further optimization errors. In this paper, a multi-objective surrogate modeling (MSM) method for telescopic boom forklifts based on closed-loop transfer learning is proposed in order to solve these challenges. The MSM consists of the following two steps: to pre-train an initial deep neural network model (deep model) with a large amount of existing simulation data from the same type of forklift and to transfer the model with a small amount of measurement data collected on the current forklift. A general framework for deep neural network (DNN) training is introduced to improve the approximation ability of the initial model. Moreover, a novel uncertainty-analysis-based sampling method is suggested for measurement data development, and combined with transfer learning to form a closed-loop mode to improve the stability of the final model. The superiority of MSM is demonstrated through comparative studies with the fine-tuning method on a telescopic boom forklift with two objectives. The experimental results show that the Correlation coefficient (R) of the deep model can reach 0.9971 by using only 80 sets of training data. In addition, it can also achieve an improvement of at least a 13.25% reduction in Root Mean Squared Error (RMSE) and a 9.19% reduction on average in Maximum Absolute Error (MAE), as well as stronger robustness compared to the benchmarks. Furthermore, it will provide a valuable reference for the simulation optimization of complex electromechanical products.

Filter pruning is necessary to efficiently deploy convolutional neural networks on edge devices that have limited computational resources and power budgets. With conventional filter pruning techniques, the same pruning rate is manually specified for different convolutional layers, which is suboptimal and time-consuming. To extract the features from the coarse level to the fine level, the number of filters in each layer has various distributions. Therefore, it is unsuitable to utilize the same pruning rate for different functional layers. To address this issue, we propose a high-dimensional Bayesian optimization-based filter pruning (HDBOFP) algorithm, which aims to automatically determine the most appropriate pruning rate for each convolutional layer. In addition, the proposed method can automatically identify optimal pruning-rate combinations without a time-consuming retraining phase. Compared with conventional filter pruning methods, this automated filter pruning technique exhibits a higher efficiency, which improves accuracy and reduces the required human labor. The effectiveness of our automated filter pruning algorithm is validated through two major computer vision applications, namely image classification and object detection. Specifically, when used in combination with ResNet-110 to classify the CIFAR-10 dataset, HDBOFP reduces the required number of float point operations (FLOPs) by more than 62% without affecting accuracy. Similarly, when HDBOFP is added to the YOLOv5l framework to run detection experiments on the MS-COCO 2017 dataset, FLOPs decrease by more than 43% with only a 1.2% loss in mean average precision, which has advanced the previous studies.

IoT devices handle a large amount of information including sensitive information pertaining to the deployed application. Such a scenario, makes IoT devices susceptible to various attacks. In addition to securing IoT devices, it is equally important to secure communication among devices and with the outside world. RS232 is a common communication protocol used in IoT and embedded devices. Hence ensuring, Trojan detection in RS232 plays a major role in providing secured communication among edge assisted IoT devices. The inclusion of malicious circuits known as hardware Trojans can occur at any stage of the IC design and manufacturing. Existing pre-silicon detection schemes with static features is limited by the number of features that are learned by the detection scheme. In contrast, machine learning allows enhanced Trojan space exploration. Existing machine learning-based Trojan detection consists primarily of supervised algorithms that rely on high-quality labeled datasets for efficient Trojan detection. Unsupervised methods, on the other hand, underperform due to limited training data and severe imbalance within the available data. To handle such a situation, a semi-supervised hardware Trojan detection has been proposed. In this work, permutation importance guided principal component analysis, correlation aware data augmentation, and hyper-parameter optimization using genetic algorithm aid in optimal dataset and model generation. Pseudo label generation using semi-supervised schemes is utilized to handle partially labeled datasets. For the Trust-HUB benchmarks, the proposed methodology achieves an average of 88.48% true positive rate and 95.77% true negative rate which, clearly indicates the effectiveness and feasibility of semi-supervised hardware Trojan detection.

This paper proposes a collaborative machine learning-based framework to detect cyber-attacks in a power system, leading to deviation in the state variable behavior. Based on the proposed architecture, three different machine learning-based methods, i.e., visualization, classification, and clustering, are employed and compared to find the best one in the FDIA detection process. To this end, pre-processing is employed in the first stage. In the second stage, the patterns of the state vectors are transferred into features. Hence, 24 statistical features, including measures of central tendency, variability, measures of shape, and position, are extracted to find various properties. Then, in the third stage, a supervised algorithm is employed to rank and find the most crucial features in FDIA. In the fourth stage, an unsupervised dimensionality reduction technique (PCA) is applied to reduce the feature space. In the fifth and last stage, visualization, classification, and clustering-based methods are developed to detect FDIA. To simulate an attack, it is assumed that an intruder decreases or increases the state vectors at different buses with various attack parameters (i.e., 0.90, 0.95, 0.96, 0.97, 0.98, 1, 1.02, 1.03, 1.04, 1.05, and 1.10). The proposed method effectiveness is assessed on the New York Independent System Operator (NYISO) data applied to the IEEE 14-bus system. The results presented in the paper from different scenarios (i.e., phase angle ($\theta $), voltage magnitude ($V_{m}$), measurements, and multiple attacks) on a real-world dataset demonstrate that the collaborative optimized PCA-Density-based machine learning technique can detect most of the attack samples with good performance scores (i.e., recall, precision, F1) and outperforms the other investigated methods. Moreover, it is general and adaptable enough to cover the situation where either the system characteristics or the attack behavior changes. [ABSTRACT FROM AUTHOR]

Software testing is an important task in software development activities, and it requires most of the resources, namely, time, cost and effort. To minimize this fatigue, software bug prediction (SBP) models are applied to improve the software quality assurance (SQA) processes by predicting buggy components. The bug prediction models use machine learning classifiers so that bugs can be predicted in software components in some software metrics. These classifiers are characterized by some configurable parameters, called hyper-parameters that need to be optimized to ensure better performance. Many methods have been proposed by researchers to predict the defective components but these classifiers sometimes not perform well when default settings are used for machine learning classifiers. In this paper, software bug prediction model is proposed which uses machine learning classifiers in conjunction with the Artificial Immune Network (AIN) to improve bug prediction accuracy through its hyper-parameter optimization. For this purpose, seven machine learning classifiers, such as support vector machine Radial base function (SVM-RBF), K-nearest neighbor (KNN) (Minkowski metric), KNN (Euclidean metric), Naive Bayes (NB), Decision Tree (DT), Linear discriminate analysis (LDA), Random forest (RF) and adaptive boosting (AdaBoost), were used. The experiment was carried out on bug prediction dataset. The results showed that hyper-parameter optimization of machine learning classifiers, using AIN and its applications for software bug prediction, performed better than when classifiers with their default hyper-parameters were used.

The success of Convolutional Neural Networks is highly dependent on the selected architecture and the hyper-parameters. The need for the automatic design of the networks is especially important for complex architectures where the parameter space is so large that trying all possible combinations is computationally infeasible. In this study, Microcanonical Optimization algorithm which is a variant of Simulated Annealing method is used for hyper-parameter optimization and architecture selection for Convolutional Neural Networks. To the best of our knowledge, our study provides a first attempt at applying Microcanonical Optimization for this task. The networks generated by the proposed method is compared to the networks generated by Simulated Annealing method in terms of both accuracy and size using six widely-used image recognition datasets. Moreover, a performance comparison using Tree Parzen Estimator which is a Bayesion optimization-based approach is also presented. It is shown that the proposed method is able to achieve competitive classification results with the state-of-the-art architectures. When the size of the networks is also taken into account, one can see that the networks generated by Microcanonical Optimization method contain far less parameters than the state-of-the-art architectures. Therefore, the proposed method can be preferred for automatically tuning the networks especially in situations where fast training is as important as the accuracy.

Data-driven optimization is an efficient global optimization algorithm for expensive black-box functions. In this paper, we apply data-driven optimization algorithm to the task of change detection with synthetic aperture radar (SAR) images for the first time. We first propose an easy-to-implement threshold algorithm for change detection in SAR images based on data-driven optimization. Its performance has been compared with commonly used methods like generalized Kittler and Illingworth threshold algorithms (GKIT). Next, we demonstrate how to tune the hyper-parameter of a (previously available) deep belief network (DBN) for change detection using data-driven optimization. Extensive evaluations are carried out using publicly available benchmark datasets. The obtained results suggest comparatively strong performance of our optimized DBN-based change detection algorithm.

With the development of healthcare technologies, the elderly population has grown and therefore populating ageing has emerged as a social issue. It is a cause of rise in patients with geriatric disorders, among which dementia is very fatal to the elderly's activities of daily living. In the studies on dementia risk prediction, a method using deep learning was proposed. It requires a lot of image data and much time to learn. Therefore, this study proposes a prediction model of dementia risk based on XGBoost using derived variable extraction from numericalized dementia data and hyper-parameters optimization. The proposed method extracts variable importance from typical independent variables with the use of gradient boosting and then generates derived variables. The generated derived variables are applied to variable importance analysis and thereby a Top-N group is created. Then, for achieving optimal performance in line with the data characteristics of each Top-N group, hyper-parameter tuning is conducted. With the optimized groups, XGBoost model based performance is evaluated. In addition, for the performance evaluation of the proposed model, goodness-of-fit for machine learning classification models is evaluated. According to the Top-N group performance evaluation with different numbers of derived variables, Top-20 model showed the best performance, and the optimized hyper-parameter values were eta = 0.10, gamma = 0, max_depth = 4, and min_child_weight = 1. As a result, the accuracy of the XGBoost model proposed in this study was 85.61%, and its F1-score was 79.28%. When the proposed model is compared with Decision Tree, Random Forest, SVM, and k-NN models, it has the best performance.

Objective: Parkinson's disease (PD) is a serious neurodegenerative disorder. It is reported that most of PD patients have voice impairments. But these voice impairments are not perceptible to common listeners. Therefore, different machine learning methods have been developed for automated PD detection. However, these methods either lack generalization and clinically significant classification performance or face the problem of subject overlap. Methods: To overcome the problems discussed above, we attempt to develop a hybrid intelligent system that can automatically perform acoustic analysis of voice signals in order to detect PD. The proposed intelligent system uses linear discriminant analysis (LDA) for dimensionality reduction and genetic algorithm (GA) for hyperparameters optimization of neural network (NN) which is used as a predictive model. Moreover, to avoid subject overlap, we use leave one subject out (LOSO) validation. Results: The proposed method namely LDA-NN-GA is evaluated in numerical experiments on multiple types of sustained phonations data in terms of accuracy, sensitivity, specificity, and Matthew correlation coefficient. It achieves classification accuracy of 95% on training database and 100% on testing database using all the extracted features. However, as the dataset is imbalanced in terms of gender, thus, to obtain unbiased results, we eliminated the gender dependent features and obtained accuracy of 80% for training database and 82.14% for testing database, which seems to be more unbiased results. Conclusion: Compared with the previous machine learning methods, the proposed LDA-NN-GA method shows better performance and lower complexity. Clinical Impact: The experimental results suggest that the proposed automated diagnostic system has the potential to classify PD patients from healthy subjects. Additionally, in future the proposed method can also be exploited for prodromal and differential diagnosis, which are considered challenging tasks.

Unlike the existing research that focuses on detecting electricity theft cyber-attacks in the consumption domain, this paper investigates electricity thefts at the distributed generation (DG) domain. In this attack, malicious customers hack into the smart meters monitoring their renewable-based DG units and manipulate their readings to claim higher supplied energy to the grid and hence falsely overcharge the utility company. Deep machine learning is investigated to detect such a malicious behavior. We aim to answer three main questions in this paper: a) What are the cyber-attack functions that can be applied by malicious customers to the generation data in order to falsely overcharge the utility company? b) What sources of data can be used in order to detect these cyber-attacks by the utility company? c) Which deep machine learning-model should be used in order to detect these cyber-attacks? Our investigation revealed that integrating various data from the DG smart meters, meteorological reports, and SCADA metering points in the training of a deep convolutional-recurrent neural network offers the highest detection rate (99.3%) and lowest false alarm (0.22%). [ABSTRACT FROM AUTHOR]