Journal of Advances in Information Technology

Constructing IoT Botnet Detection Model Based on Degree Centrality and Path Analysis

Fri, 08 Mar 2024 00:00:00 +0800

Internet of things (IoT) Botnet is a network of connected devices, generally smart devices with software and intelligent sensors, networked over the internet to send and receive data from other intelligent devices infected with IoT Botnet malware. IoT Botnet activities is challenging task in order to identify since IoT Botnet are targeting IoT devices. In addition, the current IoT Botnet detection is still unable to reveal patterns of IoT Botnet attacks and ignore the important recognization of IoT Botnet behaviors has resulted loss of meet the detection criteria. Thus, the focus of this research is to identify IoT Botnet behaviour, propose an IoT Botnet attack pattern based on its behaviour, construct an IoT Botnet detection model and to validate the selection of the IoT Botnet detection model utilising detection of the IoT Botnet attack detection criteria. Furthermore, this research is constructing the IoT Botnet attack pattern based on combining the IoT Botnet life cycle and IoT Botnet behaviour through the IoT Botnet activities. Then, this research has developed an IoT Botnet detection model based on graph analytics approach respectively to detect IoT Botnet attack activities. The earlier detection of IoT Botnet has been visualized by IoT Botnet attack patterns using the degree centrality and path analysis. The result showed that the proposed IoT Botnets model has accomplished the detection criteria’s.

An Efficient and Low-Cost Hierarchical FANET Architecture Using the Attached Mobility Model to Create the Core Layer for Connecting UAVs and Ground Control Stations

Tho Mai Cuong, Ly T.H. Nguyen, Cuong Q. Nguyen, Le Huu Binh, Tu T. Vo — Tue, 16 Apr 2024 00:00:00 +0800

Due to the frequent movement of the unmanned aerial vehicles (UAVs) in Flying Ad-Hoc Network (FANET), increasing the packet delivery ratio (PDR) from the UAVs to the ground control station (GCS) is a major problem. In this research, we suggest a solution which is improved from standard UAV ad-hoc architecture to this challenge. The idea of the proposed solution is to create a two-layer hierarchical FANET, core and access layers. The core layer consists of several UAVs flying according to the attached mobility model to form a backbone connecting to the GCS. The access layer includes the remaining UAVs that use the core layer UAVs as gateways to transmit data to the GCS. The simulation results using OMNET++ show that our proposed method is more efficient in terms of PDR, throughput, and end-to-end-delay than typical FANET architecture

WhatsApp-based Cloud Service Chatbot Application for Emergencies or Disasters

Oscar Peña Cáceres, Anthony Tavara-Ramos, Toefilo Correa-Calle, Manuel More-More — Thu, 28 Mar 2024 00:00:00 +0800

Climate change and its effects have led to the presence of disruptive and dynamic climatic events. The Piura region of Peru is geographically the most affected by the El Niño phenomenon. The limited technological means of communication between the local authority and the community have generated decisions that are not very acceptable to the population. The purpose of the study was to develop a Chatbot focused on WhatsApp, Manychat, and Google Sheets services to facilitate the referral and consultation of relevant information in emergency or disaster situations, taking into account the particularities and limitations of the population of Piura, Peru. To achieve this objective, the following phrases were proposed: 1) define and establish the types of conversations with a logical, intuitive, and friendly approach; 2) design and implement the specific functionalities necessary for users to send and consult information through the Whatsapp platform service; and 3) evaluate and refine the Chatbot in terms of its effectiveness and acceptance during emergency or disaster situations. The findings highlight the solution as a digital alternative that improves communication and coordination during emergency or disaster situations in Piura. The Chatbot optimizes incident management and provides efficient responses to users with acceptable levels of satisfaction.

A Literature Review on Outlier Detection in Wireless Sensor Networks

Julio García, Luis Rivera, Jonny Perez — Thu, 14 Mar 2024 00:00:00 +0800

Wireless sensor networks have become an important element of technologies such as the Internet of Things due to their ability to obtain sensory data from the physical world in tracking and monitoring applications. However, such networks are susceptible to the presence of outliers mainly due to errors or failures in the sensor nodes or the presence of events that alter the reading patterns. To address this problem, many researchers have turned their efforts to the development of outlier detection techniques that achieve maximum detection rate with the highest possible efficiency, given the limited resources typical of this type of networks. In this study, 33 papers on outlier detection techniques in wireless sensor networks between 2018 and 2023 were analyzed with the aim of describing the characteristics of these techniques, their metrics and test conditions, application areas, and possible limitations. The results showed mostly hybrid, distributed, online and multivariate sensing proposals in addition to the exploitation of spatiotemporal correlations of the data. In terms of efficiency, almost all of them reported detection rates above 85% and in several cases up to 100% but in specific conditions; with application areas especially related to environmental monitoring and care. Finally, the most relevant limitations encountered include high computational complexity and high resource consumption, sensitivity to parameters, lack of scalability, and dependence on specific assumptions about data distribution.

Interpretation of Chlorophyll Indexes to Estimate the Health Status of Plants within a Polyculture Model Implemented in the Municipality of Ponedera, Atlántico (Colombia), using multispectral images

Luis Gualdron, Gonzalo Moreno, Oscar Gauldron — Tue, 05 Mar 2024 00:00:00 +0800

Multispectral surveys are tools that enable the collection of information from fertile soils using unmanned aerial vehicles equipped with cameras capable of capturing images in specific narrow ranges of visible and non-visible light spectra. The research presented in this article focuses on identifying various variables such as plant health, chlorophyll levels, and the presence of desired nutrients. These variables are assessed using the chlorophyll index (CI) through the development of an inspection and monitoring strategy, which allows for the rapid characterization of the soil under study and facilitates decision-making in agricultural activities. The strategy was implemented on a specific property in the municipality of Ponedera, Atlántico (Colombia), where a polyculture of beans, plantains, and cassava was cultivated. The study involved conducting multispectral surveys at multiple time points over a period of three months.

A Combined Approach Based on Antlion Optimizer with Particle Swarm Optimization for Enhanced Localization Performance in Wireless Sensor Networks

shwetha, Dr. Murthy SVN — Wed, 03 Jan 2024 00:00:00 +0800

Wireless sensor networks play essential role in daily life scenarios due to their wide range of applications. these networks are widely adopted in to accomplish several tasks such as smart cities, smart transportation, weather monitoring etc.These networks have limited resources and suffer from various challenges which impact their performance. Moreover, these networks collect the event information and if the location of information is not known then the data becomes meaningless. Therefore, localization is considered as the important aspect of these networks. Initially, GPS based localization was considered as solution for localization but these networks consist huge number of nodes which increases the cost of network deployment.GPS won't deliver accurate localization outcomes in an indoor environment. In dense network, manually establishing location reference for each sensor node is also a tedious task. This creates a situation where the sensor nodes must locate themselves without any specialised hardware, such as GPS, or manual configuration. Utilizing localization methods, WSNs may be deployed with reduced cost. Localization accuracy and complexity still remains the challenging issue for traditional methods. Therefore, in this work, we introduce optimization based method where we consider antlion optimization as base method and incorporate particle swarm based position and velocity update method to increase the localization performance. The experimental study shows that the average localization error is obtained as 0.06525m, 0.08125m, 0.1175m, 0.3m and 0.575m using Proposed Model, CSO, PeSOA, PSO and BPSO, respectively.

Approach of Item-Based Collaborative Filtering Recommendation Using Energy Distance

Tu Tran, Hiep Huynh, Lan Phan — Wed, 03 Jan 2024 00:00:00 +0800

The current collaborative filtering recommendation method using energy distance only focuses on the relationship between the user and the user, between the user group and the user group. This method has not considered relationship between the item and the item. In this article, we mainly focus on proposing an item-based collaborative filtering model using the energy distance. The proposed model is evaluated on two popular datasets Jester5k and MovieLens100k. Besides, this proposed model is also compared with two item-based collaborative filtering models using the Cosine and Pearson measures. The experimental results have shown that the Precision, Recall and F1 values of the proposed model are better than two compared models.

Steering Angle Prediction for Autonomous Vehicles Using Deep Transfer Learning

Hoang Tran Ngoc — Thu, 25 Jan 2024 00:00:00 +0800

Self-driving cars are poised to become the primary means of transportation for future generations due to their high reliability, safety, and continuous learning capabilities. Researchers are developing multiple autonomous driving systems using techniques such as behavioral cloning and reinforcement learning. Most of these systems operate in a similar manner, with the vehicle utilizing its previous experiences and knowledge of its surroundings to make informed decisions about future actions. The aim of this study is to develop a model that can replicate a driver's behavior by utilizing transfer learning from a pre-trained VGG19 model with various activation functions. The model proposed is trained to analyze images of the road ahead as input and predict the appropriate steering wheel angle adjustment needed. The results show that the model can be trained quickly and achieves an accuracy of the mean percentage of prediction (MPP) higher than 90,1%. To evaluate the performance of our dataset obtained from the robot operating system (ROS2) simulation environment, we compared the results of several convolutional neural networks (CNN) models, including CNN of Nvidia, MobileNet-V2, ResNet50, VGG16, and VGG19. Additionally, we investigated the impact of activation functions such as ELU, ReLU, and Leaky ReLU on the transfer learning model.

Study of Manhattan and Region growing Methods for Brain Tumor Detection

Suhendro Irianto, Nila, Dona — Mon, 05 Feb 2024 00:00:00 +0800

This paper explores the region-growing segmentation and content-based image retrieval (CBIR) techniques to predict brain cancer, particularly brain tumors. Cancer is one of the diseases that causes spasms in the human brain, which is one of the most important human organs. When it affects other organs, it becomes reactive and causes death. Children, teenagers, and adults could be prompted by cancer disease, but productive ages are considered very frequently. Lately, medical sciences have advanced at such a rapid pace that diagnosis techniques and medicine have given many patients hope that they can be cured. Recently, the most common problems in diagnosing brain cancer have been time-consuming, inconsistent, inaccurate, and costly. Therefore, this paper attempts to find an alternative way to solve the problem of cancer diagnosis by using artificial intelligence, particularly in the area of computer vision, and specifically using region-growing segmentation and CBIR methods. We used region-growing segmentation and CBIR to predict a brain tumor by examining brain CT-scan images. More than 800 images were used in the work, which was collected from Kaggle.com and a hospital in Lampung, Indonesia. The accuracy of developing area-region-growth segmentation methods was computed by using receiver operating characteristics (ROC) and examining the deterioration quality of specific regions in brain CT-scan images. The research demonstrates accuracy using segmentation methods of 79% by deploying 250 normal brain CT-scan and 250 brain cancer CT-scan images. At the same time, the accuracy of brain image retrieval reaches more than 96% and 94% using Manhattan as well as Euclidean distance metrics, respectively.

Evaluation of Multipath Based Protocols in Wireless Networks

Pushpender Sarao — Mon, 05 Feb 2024 00:00:00 +0800

In wireless communications, it is very tedious task to achieve best performance results without any data and time loss. It is also a challenging work to select an appropriate routing protocol in different network scenarios keeping in mind the several performance parameters. In this research work reactive and proactive routing protocols are evaluated in respect of their efficiency in seven network scenarios by considering six performance parameters. Network scenarios are generated by varying the pause time, network size, simulation time, speed, number of nodes, network connections, and packet size. Performance parameters are considered as throughput, packet delivery ratio, average end to end delay, total dropped packets, jitter, and total received packets. Experimental work is conducted on network simulator NS-2.35 and results are tabulated for analytical discussion purpose. In most of the cases simulation results clearly indicate that, AOMDV and DSDV routing protocols shows excellence performance.

Improved Encryption Algorithm for Public Wireless Network

Christopher Khosa, Topside Mathonsi — Fri, 16 Feb 2024 00:00:00 +0800

Security mechanisms such as cryptographic algorithms play an important role in Wireless Fidelity (Wi-Fi). However, these algorithms consume a lot of memory and power. In this paper, we proposed an improved encryption algorithm that reduces the heavy consumption of power and storage to effectively protect public Wi-Fi networks. The proposed encryption algorithm was based on a hash-based message authentication algorithm. The proposed cryptographic algorithm was evaluated against the current cryptographic algorithm. We used the Network Simulation 2 (NS-2) tool to evaluate different settings for each algorithm, such as data block size, different platforms, different and decryption speeds.

Minimizing the Energy Consumption and Exploiting the NLT by E2HCA Model in WSN

Saritha Siddamsetty, Sreenivasa Reddy Edara — Fri, 16 Feb 2024 00:00:00 +0800

In fresh years, Wireless Sensor Network (WSN) has arose as a practical option for many sectors in need of smart technology. Despite its impressive credentials, WSN's excessive need for power remains a significant limitation. There is a pressing need to create a trustworthy WSN with efficient energy and network lifetime due to the proliferation of tiny sensors with incomplete resources. One of the actual ways to deal with these matters is to divide the nodes into clusters. So, it is crucial to make efficient use of existing energy to prevent energy waste. In this study, we suggest an E2HCA to reduce the high energy ingesting and increase the network lifetime of WSNs. The position and speed of a gas molecule, among other KGMO particle characteristics, are initially determined by calculating their kinetic energy. Hybridization of KGMO arises from KGMO's faster convergence in space. Pelican Optimization Algorithm (POA) is used to fix the problems with KGMO by changing its inertia weight. Throughput, the number of live/dead nodes are all validated against other prominent meta-heuristic methodologies in a MATLAB environment simulation of the proposed KGMO-POA.

Enhancement of Recommendation Engine Technique for Bug System Fixes

Jalal S. Hameed , Mohammed Al-Shammaa , fuart tawfeeq — Sun, 28 Apr 2024 00:00:00 +0800

This study aims to develop a recommendation engine methodology to enhance the model's effectiveness and efficiency. The proposed model is commonly used to assign or propose a limited number of developers with the required skills and expertise to address and resolve a bug report. Managing collections within bug repositories is the responsibility of software engineers in addressing specific defects. Identifying the optimal allocation of personnel to activities is challenging when dealing with software defects, which necessitates a substantial workforce of developers. Analyzing new scientific methodologies to enhance comprehension of the results is the purpose of this analysis. Additionally, developer priorities were discussed, especially their utility in allocating a problem to a specific developer. An analysis was conducted on two key areas: first, the development of a model to represent developer prioritizing within the bug repository, and second, the use of hybrid machine learning techniques to select bug reports. Moreover, we use our model to facilitate developer assignment responsibilities. Moreover, we considered the developers' backgrounds and drew upon their established knowledge and experience when formulating the pertinent objectives. An examination of two individuals' experiences with software defects and how their actions impacted their rankings as developers in a software project is presented in this study. Researchers are implementing developer categorization techniques, assessing severity, and reopening bugs. A suitable number of bug reports is used to examine the model's output. A developer's bug assignment employee has been established, enabling the program to successfully address software maintenance issues with the highest accuracy of 78.38%. After a brief introduction, the article discusses the background and methodology process in Section 3. A detailed description of the methodology used in this study is presented in Section 4 of this work. In section 5, clustering is used to make the expected outcomes. The final section of this study discussed the study's conclusion and prospects.

Efficient Brain Tumor Classification with a Hybrid CNN-SVM Approach in MRI

Shweta Suryawanshi, Dr. Sanjay B. Patil. — Fri, 08 Mar 2024 00:00:00 +0800

Brain Magnetic Resonance Imaging (MRI) is a crucial diagnostic tool in neuroimaging that provides valuable insights into various neurological disorders. Accurate classification of brain MRI images is vital in aiding medical professionals in diagnosis and treatment planning. The multiclass classification of brain MRI images has significant implications in clinical practice. Accurate classification can aid in detecting and characterizing various brain abnormalities, including tumors, haemorrhages, and neurological disorders. Our suggested strategy can help doctors make prompt and accurate diagnoses by automating the classification process and improving patient care and results. This study uses the two common datasets, brats and Sartaj, to propose a thorough method for multiclass classification of brain MRI utilizing CNN, VGG19, and the CNN-SVM algorithm. The proposed approach leverages the power of deep learning for feature extraction and the versatility of Support Vector Machines (SVM) for classification. Firstly, a CNN model is trained to extract discriminative features from brain MRI images. The VGG19 architecture, a widely used pre-trained CNN, is employed as a feature extractor. By utilizing the pre-trained weights of VGG19, the model can effectively capture high-level representations of the input images. The results demonstrate the efficacy of this method in accurately classifying brain MRI images. Further research can explore the application of this approach in larger datasets and investigate other deep learning architectures for feature extraction, providing further advancements in medical image analysis and diagnosis.

Improving System Accuracy by Modifying The Transfer Learning Architecture For Detecting Clove Maturity Levels

Firman Tempola, Rosihan — Tue, 26 Mar 2024 00:00:00 +0800

Detecting the maturity level of cloves is the initial stage in getting quality cloves. Early recognition of the maturity level of cloves is an essential stage in the clove industry. The maturity level of clove flowers can provide valuable information to clove farmers regarding clove harvest time. During the process of determining the level of maturity, it still relies on visual observation. This causes novice farmers and clove workers to still make mistakes in determining the start of the clove harvest. For this reason, in this research, initial detection of the maturity level of cloves was carried out based on images of clove flowers. There are four maturity levels: mature cloves, semi-ripe cloves, over-ripe cloves, and dry cloves. The proposed research method is a modification of the transfer learning architecture. The research results show that modifying the Transfer learning architecture by adding three layers can increase system accuracy in the VGG16 and ResNet50 models. Meanwhile, for the VGG19 model, accuracy increased only when initializing the number of epochs to 10

Object Detection By Effective Segmentation of Tree Canopy using U-Net Model

Vasavi S, Lakshmi Likhitha Atluri, Sai Premchand Veeranki, Yasaswini Jampa — Tue, 26 Mar 2024 00:00:00 +0800

According to the Forest Survey of India and yearly report, Kerala has a total area covered by trees of 2951 sq km, which is 7.59% of the state's total area. In regions like Kerala, identifying objects from Very High Resolution Satellite (VHRS) images has become a major challenge. Using a method known as “tree canopy”, which refers to the area shaded by trees, objects that are covered by trees can be identified. For mapping tree crowns Mask R-CNN is used. It struggles for accurate segmentation and distinguish objects from the surrounding trees, leading to misclassification and incorrect object masks. It can also be computationally expensive, making it challenging to process high-resolution images in real-time. A deep learning model that uses a semantic segmentation approach is proposed to detect tree canopy covering objects. Dataset with 0.5m resolution images is prepared from SAS Planet images. The input image is now preprocessed using pre-processing techniques and trained with U-Net. Further, the images are closed using morphological operation to detect the object. The model is evaluated for 25 epochs with an accuracy of 92%. Finally, the objects are classified based on semantic segmentation using U-Net with backbone of ResNet34. The objects are classified as buildings, roads, water bodies and got accuracy of 84 %.

Binary Classification of Heart Disease based on Differential Evolution-Optimised Machine Learning Approach

Theodore Nicholas Richard Egling — Tue, 09 Apr 2024 00:00:00 +0800

Accurate and timely diagnosis of heart disease is a persistent challenge in healthcare, necessitating innovative diagnostic methodologies. This study investigates the efficacy of Differential Evolution (DE) for hyperparameter optimisation in machine learning algorithms, targeting improved performance in heart disease binary classification. DE was selected for its robustness and ability to efficiently navigate high-dimensional parameter spaces, essential attributes for the fine-tuning of complex models. Employing the Cleveland Heart Disease dataset, the study optimised three machine learning classifiers: RandomForest, AdaBoost, and Gradient Boosting. Post-optimisation, the DE-enhanced RandomForest Classifier achieved a standout performance with an accuracy of 93.3% and an F1-score of 90.9%. Likewise, AdaBoost and Gradient Boosting classifiers also exhibited performance gains, reaching accuracies of 88.9% and 86.7%, and F1-scores of 85.7% and 83.3%, respectively. These results not only outperform various existing models but also offer insights into the differential impacts of DE on multiple algorithms. The study lays a solid foundation for future research and clinical applications, indicating that DE-optimised machine learning algorithms hold significant promise for advancements in cardiovascular disease diagnostics.

Deep Learning-Based Lane-Keeping Assist System for Self-Driving Cars Using Transfer Learning and Fine Tuning

Fri, 08 Mar 2024 00:00:00 +0800

This paper presents an advanced lane-keeping assistance system specifically designed for self-driving cars. The proposed model combines the powerful Xception network with transfer learning and fine-tuning techniques to accurately predict the steering angle. By analyzing camera-captured images, the model effectively learns from human driving knowledge and provides precise estimations of the steering angle necessary for safe lane-keeping. The transfer learning technique allows the model to leverage the extensive knowledge acquired from the ImageNet dataset, while the fine-tuning technique is utilized to tailor the pre-trained model to the specific task of steering angle prediction based on input images, enabling optimal performance. To evaluate the system's effectiveness, a comprehensive comparative analysis is conducted against popular existing models, including Nvidia, MobilenetV2, VGG19, and InceptionV3. The evaluation includes an assessment of the operational accuracy based on the loss function, specifically utilizing the Mean Squared Error (MSE) equation. The proposed model achieves the lowest loss function values for both training and validation, demonstrating its superior predictive performance. Additionally, the model's performance is further evaluated through extensive real-world testing on pre-designed trajectories and maps, resulting in the minimal deviation of the steering angle from the desired trajectory over time. This practical evaluation provides valuable insights into the model's reliability and its potential to effectively assist in lane-keeping tasks.

Assamese dialect identification using static and dynamic features from vowel sound

Hem Das, Utpal Bhattacharjee — Tue, 27 Feb 2024 00:00:00 +0800

This paper presents a novel approach to Assamese dialect identification by leveraging the acoustic and prosodic behavior of speech signals derived from vowel sounds. The distinctive nature of the dialect signals is captured by employing acoustic characteristics such as formants (F1, F2, and F3) and prosodic characteristics including energy, fundamental frequency (F0), and duration. A comprehensive vowel speech corpus is collected, comprising recordings from native Assamese speakers representing four different dialectal regions, to assess the proposed methodology. To extract relevant features, frame-level statistical features are derived from vowel sounds, while temporal dynamic features are extracted from steady-state vowel regions. For data collection, a phonetically rich script is utilized to record speech from four dialectal varieties for a read speech database, and partly spontaneous speech is recorded from interactions between speakers representing the four dialects. Several classification methods, including three decision tree-based classifiers (RF, ERF, and XGB), are employed to discover four distinct dialects. The performance of each feature, both static and dynamic, is assessed individually. The findings demonstrate that the identification of Assamese dialects is significantly affected by several factors, namely the length of speech, its intensity, the pitch at which it is spoken, and the frequencies of the three formants. In order to determine the significance of individual features in distinguishing dialects and to evaluate the combined impact of all features on the dialect identification system, single-factor ANOVA tests were conducted. Significantly, when static features are utilized in combination with the ERF (Extreme Random Forest) ensemble model, the overall performance of dialect identification reaches 77%. The effectiveness of the suggested approach in accurately classifying Assamese dialects is demonstrated, underscoring the pivotal role of acoustic and prosodic features. To summarize, this research paper presents a strong framework for identifying Assamese dialects, demonstrating the capacity of acoustic and prosodic characteristics to capture the subtle variations within dialects. The findings of this study enhance our comprehension of dialect discrimination and lay the groundwork for the advancement of more sophisticated systems for identifying dialects.

Efficient MLTL Calibration Model for Monitoring the Real-Time Pollutant Emission from Brick Kiln Industry

Sahaya Sakila Varghese, Manohar S — Sun, 28 Apr 2024 00:00:00 +0800

Coal-ablaze Brick Kiln industries are the major contributors of Particulate Matter (PM_2.5, PM₁₀) emissions that endanger the environment and pose a variety of health risks to all the living beings. Current static ambient pollutant monitoring stations are limited to their expensive deployment. Recent advancements in Internet of Things (IoT) technology tends to have portable sensors to monitor the quality of air. Calibration for these portable sensors requires training data from static reference monitoring stations. In this study, Brick Kiln industry, which are usually remotely located from the reference stations, is chosen to monitor its emission through the IoT devices, and the calibration for the portable sensors are performed using data from a reference sensor. Calibration of the sensor reading is performed using proposed Meta Learning based Transfer Learning (MLTL) and its performance is evaluated utilizing evaluation metrics of various Machine Learning (ML) and Deep Learning (DL) based regression models. The proposed model shows the most significant scores 0.992236, 0.0002, 0.0048 for the evaluation metrics, R-squared, RMSE, and MAE respectively, as compared to other ML models while calibrating the PM pollutant’s emission rate obtained from the industry.

Using SMOTE Upsampling to Enhance the Accuracy of Sentiment Analysis for Borobudur Temple Visitor Review

Candra Agustina, Purwanto Purwanto, Farikhin Farikhin — Tue, 09 Apr 2024 00:00:00 +0800

The level of satisfaction of visitors to tourist destinations can be known from reviews on social media. One method used is to carry out sentiment analysis of comments given by visitors on social media or related websites. This study is envisioned as a preliminary phase to bolster subsequent research concerning tourist destination recommendation systems around the Borobudur Temple. We conducted sentiment analysis using a semi-supervised learning approach. Within this approach, the dataset is partitioned into labeled and unlabeled data. The labeled data serves as a reference for the automatic labeling process, which utilizes the Multinomial Naïve Bayes algorithm. Specifically, the objective is to extract sentiments from visitors to the Borobudur Temple. These extracted sentiments will later be employed as a variable in subsequent research. Dataset preprocessing steps encompass data cleaning, sentence segmentation, tokenization, and stop word removal. We observed that the difference in labeling outcomes between datasets trained without SMOTE Upsampling and those trained with SMOTE Upsampling was a mere 0.18%. The labeled data is utilized not only for training the model but also for gauging the accuracy of the Multinomial Naïve Bayes algorithm. After implementing the SMOTE Upsampling technique, we achieved an accuracy of 84.51%. Our analysis indicates superior performance when the training data undergoes the SMOTE Upsampling process.

A Hybrid Feature Extraction and Feature Selection Mechanism to Predict Disease in Plant Leaves

Abisha A, Bharathi N — Tue, 09 Apr 2024 00:00:00 +0800

The health of the plants is vital to meet the demands of the food cycle. As the symptoms of disease or infection are most commonly seen in plant leaves, the selection of features from plant leaves that are highly impacting plant health is crucial. Feature extraction(FE) and Feature selection(FS) are significant in Deep learning(DL) and Machine learning(ML) models, which are used for classification and prediction. This article contributes to agriculture-based countries and benefits farmers in various ways. In this article, FE is performed using Xception pre-trained model and the extracted features are sent for FS. Further, six FS methods such as ANOVA, chi-square, Sequential Forward Selection (SFS), Sequential Backward Selection (SBS), Lasso and ridge, have been deployed and compared with machine learning algorithms such as logistic regression(LR), K Nearest Neighbours(KNN), Decision-Trees(DT), Random-Forest(RF), Support-Vector-Machine(SVM), Naive-Bayes(NB) for classification. The article also proposes an Ensemble Feature Selection(EFS)-RF method, which combines feature sets from six feature selection algorithms and classifies based on majority voting. The contribution of the paper is to benefit agriculture by using hybrid DL(FE using Xception) and ML(Classification using RF) with an ensemble of the above-mentioned FS methods and finding the best feature that is obtained in the majority of the subsets and adding higher weightage to the feature obtained from most of the models. The proposed method has outperformed other algorithms for both datasets with 98 % accuracy and 0.02 MSE for dataset I and 98.125 % accuracy and 0.01875 MSE for dataset II.

Automatic gender authentication of Arabic speech using deep learning

Amjad Rehman Khan — Wed, 24 Apr 2024 00:00:00 +0800

Speech recognition is progressively being utilized in practical applications with time. Automatic gender identification is one of the most intriguing applications since it distinguishes female and male speeches from briefly spoken communication records. This is advantageous in various applications, including automated conversation systems, system verification, demographic attribute prediction and assessing speaker’s expressions. Speech is a natural mode of communication, and the pitch variation of a gender-specific speech signal is often used to identify a person as male or female. Automatic voice recognition is a field of study that enables computers to take vocal input from people and interpret it with higher accuracy. Numerous methods exist for implementing voice recognition models. One of the developing methods for voice recognition is using neural networks with deep learning. Arabic is a widely spoken natural language that received less attention regarding voice recognition. We present a method for determining gender from the speech in this research by integrating audio preprocessing, MFCC, Delta MFCC, and Log Filter bank feature extraction. Gender categorization using a feed-forward neural network and a Keras base neural network is evaluated using the Urban Jordanian Arabic (UJA) Arabic dataset. The suggested system surpasses current methods in terms of accuracy.

Development and Comparison of Multiple Emotion Classification Models in Indonesia Text Using Machine Learning

Ahmad, Sarjon Defit, Gunadi Widi Nurcahyo — Wed, 24 Apr 2024 00:00:00 +0800

Emotions are individual responses to an event or situation. Studies have been conducted to extensively discuss emotions, including the usage of sentiment analysis to analyze digital texts in order to determine their positive or negative emotional tone. Therefore, this study was conducted using the dataset consisting of different tweets retrieved from the social media platform, Twitter, with a focus on those posted in relation to the 2024 presidential election in Indonesia. The dataset was divided into two test sets including the first with six labels such as sadness, fear, anger, surprise, love, and joy while the second consisted of positive and negative labels. Several machine learning algorithms including Naïve Bayes (multinomial Bayes, Bernoulli Bayes, Complement Bayes), K-Nearest Neighbors (KNN), and Support Vector Machines (SVM) were employed to process the dataset. Moreover, two features extraction methods including TF-IDF and BoW were utilized for comparison purposes and they both led to unsatisfactory accuracy in the SVM. This led to the inclusion of a kernel combination approach, referred to as PoRLI (Polynomial, RBF, and Linear), to improve the accuracy. The SVM PoRLi was observed to have produced a significant improvement in accuracy compared to using single kernel.

Enhancing Text Sentiment Classification with Hybrid CNN-BiLSTM Model on WhatsApp Group

susandri susandri, Sarjon Defit, Muhammad Tajuddin — Thu, 14 Mar 2024 00:00:00 +0800

Large amounts of data are generated from social media. The need to extract meaningful information from big data, classify it into different categories, and predict user sentiment is crucial. Text classification is a representative research topic in the field of natural language processing that categorizes unstructured text data into sentiments to make it more meaningful. Improving word and text category accuracy requires more precise text classification methods. Deep Learning models developed and implemented in this field have shown progress, but further improvement is still needed. This paper utilizes the NLP process on a WhatsApp group dataset to determine sentiment, testing it with five Deep Learning models: Neural Network, RNN, LSTM, Bidirectional LSTM, CNN, and proposes a hybrid CNN-BiLSTM model. The proposed model employs feature extraction and a hybrid architecture with activations, dropouts, filters, kernel sizes, and different layers to classify text sentiment. To verify the performance of the proposed model, it is compared with previous studies. In single-model testing, the LSTM and BiLSTM achieves the best accuracy of 81%. Meanwhile, the proposed model has reached an accuracy of 88% on the utilized dataset. By comparing the performance of the proposed model with previous studies, the proposed model offers better sentiment classification performance.

A Machine Learning Approach for Stroke Differential Diagnosis by Blood Biomarkers

khaled Sayed, Fayroz farouk — Wed, 03 Jan 2024 00:00:00 +0800

Stroke happens when a clot blocks the blood supply to a region of the brain (ischemic stroke) or when an artery ruptures or spills blood (hemorrhagic stroke). Seeking medical care after a stroke may increase one's chances of survival and reduce long-term brain damage. Neuroimaging helps determine who and how to treat, although it is costly, not always accessible, and may have contraindications. These constraints lead to these reperfusion treatments being underutilized. Using a blood biomarker panel capable of consistently differentiating between ischemic stroke and intracerebral hemorrhage might be very beneficial and straightforward to deploy. Therefore, this study describes a system to speed and improve stroke diagnosis. Using four machine learning algorithms: support vector machine (SVM), adaptive neuro-fuzzy inference system (ANFIS), K-nearest neighbor (KNN), and decision tree (DT), we aim to find promising blood biomarker candidates for differential stroke diagnosis. A two-stage binary classifier model was created to classify the stroke group vs. the normal group and then categorize the instances allocated to the stroke group into ischemic and hemorrhagic groups. Our findings reveal that SVM is better than ANN, ANFIS, and DT for distinguishing strokes in Egyptian patients, according to our data. The most important blood features are ABS Neutro, CPK, Neutro/Neutrophils, and WBC Count/Leukocytes laboratory tests that may serve as crucial and significant indications for stroke diagnosis. The selected characteristics and a two-stage binary classifier discriminated with 100% accuracy (Ischemic and hemorrhagic patients). This method for identifying and classifying brain strokes was accurate, easy to use, and cost-effective.

Preprocessing Strategy to Improve the Performance of Convolutional Neural Networks Applied to Steganalysis in the Spatial domain

Mario Alejandro Bravo Ortiz — Tue, 09 Jan 2024 00:00:00 +0800

Recent research has shown that deep learning techniques outperform traditional steganography and steganalysis methods, which has contributed in several researches to propose different types of increasingly complex and larger convolutional neural networks (CNNs) to detect steganographic images, which aims to outperform the state of the arts most of the time in a 1%-2%. This paper presents a data preprocessing and distribution strategy that improves accuracy and convergence during training. The strategy implements a bifurcation of spatial rich model (SRM) filters and DCT filters, which are a set on one branch as trainable and on the other untrainable, followed by three blocks of residual convolutions and an excitation layer. The proposed strategy improves the accuracy of CNNs applied to steganalysis by 2%-15% while preserving the stability.

Exploratory Architectures Analysis of Various Pre Trained Image Classification Models for Deep Learning

Gokila S, Deepa S, Loveline Zeema J — Thu, 18 Jan 2024 00:00:00 +0800

The Image classification is one of the significant applications in the area of Deep Learning (DL) with respective to various sectors. Different types of neural network architectures are available to perform the image classification and each of which produces the different accuracy. The dataset and the features used influences the outcome of the model. The research community is working towards the generalized model at least to the domain specific. On this gesture the contemporary survey of various Deep Learning models is identified using knowledge information management methods to move further to provide optimal and generalized DL model to classify images narrow down to the sector specific. The study systematically presents the different types of architecture, its variants, layers and parameters used for each version of DL model and also applications and limitations of the type of architecture. It helps the researchers to select appropriate DL architecture for specific applications.

B-DT Model: A Derivative Ensemble Method to Improve Performance of Intrusion Detection System

Amarudin, Ridi Ferdiana, Widyawan — Thu, 18 Jan 2024 00:00:00 +0800

In cyber security, system security must be prioritized. Therefore, to improve system security, a system device called an Intrusion Detection System (IDS) is needed. IDS is a system that can detect suspicious activity on a system or network. The problem with IDS is that many types of attacks appear now, making it difficult to detect them. Therefore, to overcome this problem, many IDS based on machine learning have been applied. Machine Learning has been widely adopted to improve IDS performance, but false detection occurs frequently. This paper proposes a novel method called the B-DT model, which combines the bagging technique ensemble-base and Decision Tree (DT) classifier. This model addresses the limitations of existing methods. The B-DT model was trained and evaluated on NSL-KDD and UNSW-NB15 datasets. The experimental results show that it significantly improves accuracy, precision, recall, and f1-score compared to several traditional and state-of-the-art machine learning methods. This method can work well not only on binary-class data but also on multi-class labelled data. The performance of the B-DT model can achieve an accuracy of 99,45% on the NSL-KDD dataset and 79,67% on the UNSW-NB15 dataset. The statistical evaluation shows this model has increased significantly compared to other models. These results suggest that the proposed B-DT model can effectively enhance the performance of IDS and be a promising solution for practical applications.

Face Identification Based on Active Facial Patches Using Multi-Task Cascaded Convolutional Networks

Krishnaraj M, Jeberson Retna Raj R — Thu, 18 Jan 2024 00:00:00 +0800

Face recognition technology is widely used for access control, security, identification, safeguarding, verification, timekeeping, and machine vision, etc. a new face identification algorithm referred to as Multi-Task Cascaded Convolutional Network (MTCCN) has emerged and has been widely used in high accuracy and efficiency in facial recognition, active facial patch identification framework face detection, selection of eyes, nose, lip, and eyebrow, identifying facial patches location and extraction of patches. This paper aims to discuss the recognition and identification of faces using layers of the Convolutional Neural Network (CNN). This is done to process camera frames as they appear and subsequent identification of the person. With three convolutional networks, MTCCN outperforms many face detection tests incredibly well, even though it maintains real-time performance. A proposed method for recognizing human faces in real time was developed, evaluated, and 97.62 % of the time, the technique could recognize human faces correctly.

An Improved Fake News Detection Model Using Hybrid Time Frequency-Inverse Document Frequency for Feature Extraction and AdaBoost Ensemble Model as a Classifier

Lakshmi Holla, KS Kavitha — Thu, 15 Feb 2024 00:00:00 +0800

In recent times, fake news content on the internet has become one of the challenging issues which may create an impact on society and individuals. Moreover, the spread of fake news on social platforms enhances the risk of loss of trustworthiness and disseminates fake information through various platforms on the internet. So, detecting fake news which evolves on the internet plays a significant role among society and individual persons. By knowing that, this research proposed an improved fake news detection model which utilized the proposed Hybrid Time-Frequency- Inverse Document Frequency (TF-IDF) to extract the features, and classification is performed using the Adaptive boosting ensemble classifier which is a combination of Iterative Dichotomiser 3 (ID-3), Random Forest (RF) and Naïve Bayes (NB) classifiers. The Least Absolute Shrinkage and Selection Operator (LASSO) is used for selecting the features and the Adaboost ensemble classifier is used to categorize the news as fake or true. The obtained results exhibit that TF-IDF with Adaboost ensemble classifier has achieved a better classification accuracy of 98.98% which is comparatively higher than the existing N-Gram with TF-IDF and Bidirectional Encoder Representation from Transformer (BERT) and Word2Vec with Convolutional Neural Network – Bidirectional Long Short Term Memory (CNN-Bi LSTM) with 96.81% and 97.74% respectively.

Hybrid Deep Learning Network Intrusion Detection System Based on Convolutional Neural Network and Bidirectional Long Short-Term Memory

Anindra Ageng Jihado, Abba Suganda Girsang — Fri, 16 Feb 2024 00:00:00 +0800

Network security has become crucial in an era where information and data are valuable assets. An effective network intrusion detection system (NIDS) is required to protect sensitive data and information from cyberattacks. Numerous studies have created NIDS using machine learning algorithms and network datasets that do not accurately reflect actual network data flows. Increasing hardware capabilities and the ability to process big data have made deep learning the preferred method for developing NIDS. This study develops a NIDS model using two deep learning algorithms: convolutional neural network (CNN) and bidirectional long-term memory (BiLSTM). CNN extracts spatial features in the proposed model, while BiLSTM extracts temporal features. Two publicly available benchmark datasets, CICIDS2017 and UNSW-NB15, are used to evaluate the model. The proposed model surpasses the previous method in terms of accuracy, achieving 99.83% and 99.81% for binary and multiclass classification on the CICIDS2017 dataset. On the UNSW-NB15 dataset, the model achieves accuracies of 94.22% and 82.91% for binary and multiclass classification, respectively. Moreover, principal component analysis (PCA) is also used for feature engineering to improve the speed of model training and reduce existing features to ten dimensions without significantly impacting the model's performance.

Ovarian Tumors Detection and Classification from Ultrasound Images Based on YOLOv8

Van-Hung Le, Thi-Loan Pham — Sun, 25 Feb 2024 00:00:00 +0800

Detecting and diagnosing cancer early can increase survival rates for patients by up to 50%. Ovarian cancer is the 8th leading cause of death in women. Therefore, the problem of detecting ovarian cancer early and on common data such as ovarian ultrasound images is an issue that needs to be studied. In this paper, we perform a comparative study on the latest version of the YOLO family, YOLOv8 (YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x) for the detection and classification of ovarian tumors on OTU 2D subset (OTU 2D original set - OTU 2DOS) ovarian ultrasound images. This study performed finetuned the model to detect and classify 1 label (with or without ovarian tumor on image), 8 labels according to OTU original data, and 2 labels (benign ovarian tumor and malignant). Results of detection and classification of ovarian tumors: 1 label (P=91.26%, R=83.3%, mAP50=92.67%), 8 labels (P=71.5 %, R=68.7%, mAP50=70.6%), 2 labels (P=74.9%, R=59%, mAP50=66.8%). The accuracy of YOLOv8X is the best and higher than YOLOv7 is 19%. The calculation time of YOLOv8 is also shown, the processing time of YOLOv8x is slower than YOLOv7 (YOLOv8x is 186fps on GPU, 1.84fps on CPU). Experiments and results are presented and available.

An Integrated Deep Learning Model for Concurrent Speech Dereverberation and Denoising

Vijay Mane, Seema Arote — Sun, 25 Feb 2024 00:00:00 +0800

Speech is most likely the simplest and efficient type of human-human communication, as well as the most intuitive and effective way of human-machine interaction. Human voice is often damaged in real-world contexts by both reverberation and noise from the surroundings, which has a detrimental impact on speech intelligibility and quality. In terms of denoising, a model-based approach has been thoroughly researched, and several practical solutions have been created. In comparison, study on dereverberation has been sparse. Significant advances have been achieved in the study of a model-based strategy for dereverberation. The resultant approach may be used to any deep neural network that provides masks in the time-frequency domain with just a few extra variables that can be trained and an overhead of computation that is low for state-of-the-art neural networks. A deep learning-based approach in this article is developed that eliminates early reverberations, late reverberations, and noise from speech signals in order to enhance speech signal quality. The method is tested for simulated room impulse response data from a conference room, a seminar hall, and a room from reference paper [7] with a reverberation period of 0.3s and a variety of signal-to-noise ratio values. The proposed technique outperforms baseline multichannel dereverberation and denoising algorithms as well as a cutting-edge multichannel dereverberation and denoising algorithm, resulting in a considerable improvement.

Criminal Court Judgment Prediction System Built on Modified Bert Models

Shannen Latisha, Sean Favian, Derwin Suhartono — Sun, 25 Feb 2024 00:00:00 +0800

The high crime rate in Indonesia that occurs from year to year causes a high number of cases that must be examined, tried, and decided through the courts as stipulated in Law No. 48 of 2009 concerning Judicial Power. Therefore, this study was conducted to build a system for predicting sentences resulting from criminal court decisions in the Republic of Indonesia which is expected to facilitate the implementation of jurisprudence. The prediction system was built by comparing 6 BERT (Bidirectional Encoder Representations from Transformers) models and a RoBERTa (A Robustly Optimized BERT Pretraining Approach) model on 3 different proposed architectures: BERT Base, Hierarchical BERT + Mean Pooling, and Hierarchical BERT + LSTM (Long Short-Term Memory). The compared models include indobert-base-p1, indobert-base-uncased, legal-indobert-indonlu, legal-indobert-indolem, indobert-large-p1, indonesian-roberta-base. Those models are also compared with SVM + TF-IDF (Support Vector Machine + Term Frequency–Inverse Document Frequency) as a baseline. The legal-indobert-indolem model with the Hierarchical BERT + Mean Pooling architecture succeeded in performing multi-class classification tasks into 14 classes with the highest F1-score value of 79,8888%. Thus, the successfully created model can be further used in assisting jurisprudence as it has developed the ability to predict criminal court decisions based on similar previously documented cases.

Hand Gesture Recognition Based on Electromyography Signals and Deep Learning Techniques

Mai Hassan — Mon, 26 Feb 2024 00:00:00 +0800

Hand gesture recognition based on electromyography (EMG) signals is a challenging approach for developing natural and intuitive human-computer interfaces. In this paper, a hand gesture recognition system will be proposed that uses deep learning techniques, specifically a convolutional neural network (CNN) and a long short-term memory (LSTM) by merging them into one architecture, which named CNN+LSTM model. The CNN is used to extract relevant features from the EMG signals, while the LSTM model is used to capture the temporal dynamics of the gestures. The proposed system was trained and evaluated on two datasets publicly available. The first one, DualMyo, which includes EMG signals recorded from one subject performing 8 different hand gestures, each class of gestures has been recorded 110 times. The second dataset recorded from 36 subjects performing 8 different hand gestures. Results show that our proposed system achieved performance, with an average recognition accuracy for both data sets of about 99% for the DualMyo and about 97% for the second. To tackle the testing time issue, a second model is introduced. By adding cascading CNN and max pooling layers, we achieve a substantial reduction rate of 1/20 compared to the first model in testing time without compromising recognition accuracy significantly. Experimental results demonstrate the efficacy of this approach, making it suitable for real-time applications in gesture-controlled systems.

A Novel IOT-Based Smart and Security System Model for Large Scale Farm Sustainability

Thu, 14 Mar 2024 00:00:00 +0800

The security and automation of work for large scale farms became major problem. Across the globe small scale farms have many techniques to implements. Smart and Security (SAS) is the most important measure for all the farms to grow substantially. In this paper, the idea of involving internet of things (IOT), GSM Module, GPS modules provide security and updated information of the specification to the farmer. Fingerprint based entry system into the farm helps to improve the security. The farmer can monitor the entire farm land with the help of internet in hand and the important updates like soil moisture, temperature are sent to mobile via GSM module for every hour. The entire system is controlled using microcontroller. GPS is one more device used for the vehicles which carry the cultivated product can be lively monitored which helps in improving the security. The smart system designed in this paper will be helpful for the farmers to collect live data of stock, restriction of entry into farm for unknown persons, monitoring of field with various sensors. The entire data collected and is updated in the cloud server. A dataset is created in the cloud server and will be further utilized for other farms. The data prediction is performed used convolutional neural networks. The dataset which is created to further processed using Matlab software tool.

A Smart System for the University Chemical Laboratory using IoT

Soha Shaban — Thu, 18 Jan 2024 00:00:00 +0800

University chemical laboratories are exposed to numerous kinds of risks, including frequent explosions and accidents, theft of secret information, high-value equipment, or dual-use chemicals that might be utilized in weapons, and excessive energy use. Therefore, this paper seeks to develop an automation system for the chemical laboratory while also providing it with safety and security by utilizing IoT concepts. The proposed system is divided into three parts. The first part is security, which employs a PIR sensor and RFID module to keep the laboratory secure at all times. The second part is automation, which uses the Google Assistant embedded into the Android smartphone, HC-05 Bluetooth module, an Arduino Uno, and a relay module. It is intended to allow the laboratory supervisor to remotely monitor and control different laboratory electrical appliances. The third part is safety, which monitors the laboratory environment using a flame sensor, MQ3 smoke sensor, DHT22 temperature and humidity sensor, and MQ135 air quality sensor. The data from the sensors is transferred to the Arduino Uno, which then takes the necessary action in the event of an emergency. The results indicated that the proposed system achieved a high accuracy of up to 98%, which demonstrates that it is trustworthy, dependable, and efficient.