Machine Learning-based Detection Methods Research Articles

Defect Detection of Cell Phone Screen Using a Faster Regional Convolutional Neural Network with Multi-head Attention Mechanism

Since the glass outer screen of a cell phone is the main sensory part of the human eye when using a cell phone, the advantages and disadvantages of the cell phone screen directly affect people's sense of use. Therefore, the defect detection requirements for cell phone screens are high and need to meet the needs of high-volume factory inspection. Most of the traditional defect detection methods use visual methods, the detection results are overly dependent on the subjectivity and experience of workers, the efficiency of this method is low, and the accuracy is poor. Currently, machine learning-based detection methods are applied in numerous industries. In this paper, a faster Regional Convolutional Neural Network (R-CNN) with multi-head attention mechanism for defect detection of cell phone screen is proposed. To enhance the network's capability in extracting feature information, a four-head attention mechanism is added to the last convolutional layer of the ResNet50 network. An improved Region of Interest (ROI) Align is proposed to replace the original ROI Pooling to reduce the localization error of cell phone screen defects. Replace the original Rectified Linear Unit (ReLU) activation function with the Copy Exponential Linear Unit (CELU) activation function to expedite the convergence capability of the network. Finally, by comparing with other classical model training, the evaluation results indicate that the proposed method achieved an average accuracy of 95.71%, which is a 5.34% improvement compared to the original faster R-CNN network.

Galaxy merger challenge: A comparison study between machine learning-based detection methods

Various galaxy merger detection methods have been applied to diverse datasets. However, it is difficult to understand how they compare. Our aim is to benchmark the relative performance of merger detection methods based on machine learning (ML). We explore six leading ML methods using three main datasets. The first dataset consists of mock observations from the IllustrisTNG simulations, which acts as the training data and allows us to quantify the performance metrics of the detection methods. The second dataset consists of mock observations from the Horizon-AGN simulations, introduced to evaluate the performance of classifiers trained on different, but comparable data to those employed for training. The third dataset is composed of real observations from the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP) survey. We also compare mergers and non-mergers detected by the different methods with a subset of HSC-SSP visually identified galaxies. For the simplest binary classification task (i.e. mergers vs. non-mergers), all six methods perform reasonably well in the domain of the training data. At the lowest redshift explored $0.1<z<0.3$, precision and recall generally range between sim 70<!PCT!> and 80<!PCT!>, both of which decrease with increasing $z$ as expected (by sim 5<!PCT!> for precision and sim 10<!PCT!> for recall at the highest $z$ explored $0.76<z<1.0$). When transferred to a different domain, the precision of all classifiers is only slightly reduced, but the recall is significantly worse (by sim 20-40<!PCT!> depending on the method). Zoobot offers the best overall performance in terms of precision and F1 score. When applied to real HSC observations, different methods agree well with visual labels of clear mergers, but can differ by more than an order of magnitude in predicting the overall fraction of major mergers. For the more challenging multi-class classification task to distinguish between pre-mergers, ongoing-mergers, and post-mergers, none of the methods in their current set-ups offer good performance, which could be partly due to the limitations in resolution and the depth of the data. In particular, ongoing-mergers and post-mergers are much more difficult to classify than pre-mergers. With the advent of better quality data (e.g. from JWST and Euclid), it is of great importance to improve our ability to detect mergers and distinguish between merger stages.

EVALUATING STOCHASTIC EPIGENETIC MUTATIONS: RELIABILITY, SUBTYPES, AND AN ALTERNATIVE DETECTION APPROACH

Abstract Stochastic Epigenetic Mutations (SEMs) have emerged as promising biomarkers for aging and age-related disorders. Conceptually, a SEM is defined as an outlier in DNA methylation (DNAme) value at a specific genomic site, compared to a given population or cohort. These mutations are seldom shared among subjects, emphasizing their potential to capture the heterogeneity of age-associated DNAme changes, and their cumulative count in individuals rises with age. We hypothesized that the original statistical method for SEM detection is not well suited for DNAme data and demonstrated its reliability constraints at both the individual and aggregate levels, using technical replicates from whole blood samples. In response to these challenges, we explored the factors influencing individual SEM reliability and identified best practices for SEM detection. We also developed an alternative machine learning-based detection method. Importantly, our classifier demonstrated improvements in the proportion of shared SEM and ICC scores when tested in an independent dataset with technical replicates, maintaining the associations of higher SEM loads with age, mortality, and cardiovascular disorders in the FHS study cohort. To support the wider application of our findings and address the need for SEM analysis tools, we have introduced a novel R package, SEMdetectR. Leveraging parallel programming, this package offers rapid SEM identification using both the original and new methods, accompanied by robust downstream analysis features.

IoT Botnet Detection Using Autoencoders and Decision Trees

The use of IoT devices has grown rapidly, leading to an increase in cyber attacks that pose greater security and privacy threats than ever before. One such threat is botnet attacks on IoT devices. An IoT botnet is a group of Internet-connected IoT devices infected with malware and remotely controlled by an attacker. Machine learning techniques can be employed to detect botnet attacks. The use of machine learning-based detection methods has been shown to be effective in identifying cyber attacks. The performance of the detection system in machine learning can be improved by utilizing data reduction methods. The data reduction process in classification is used to overcome the problem of scalability and computation resources in the IoT. This paper proposes a detection system using the Autoencoder reduction method and the Decision tree classification method. The test results demonstrate that the Deep Autoencoder algorithm can reduce data and memory usage from 1.62 GB to 75.9 MB, while also improving the performance of decision tree classification, resulting in a high level of accuracy up to 100%. The Autoencoder approach in conjunction with the Decision Tree exhibits superior capabilities compared to previous studies.

MM-ConvBERT-LMS: Detecting Malicious Web Pages via Multi-Modal Learning and Pre-Trained Model

In recent years, the number of malicious web pages has increased dramatically, posing a great challenge to network security. While current machine learning-based detection methods have emerged as a promising alternative to traditional detection techniques. However, these methods are commonly based on single-modal features or simple stacking of classifiers built on various features. As a result, these techniques are not capable of effectively fusing features from different modalities, ultimately limiting the detection effectiveness. To address this limitation, we propose a malicious web page detection method based on multi-modal learning and pre-trained models. First, in the input stage, the raw URL and HTML tag sequences of web pages are used as input features. To help the subsequent model learn the relationship between the two modalities and avoid information confusion, modal-type encoding, and positional encoding are introduced. Next, a single-stream neural network based on the ConvBERT pre-trained model is used as the backbone classifier, and it learns the representation of multi-modal features through fine-tuning. For the output part of the model, a linear layer based on large margin softmax is applied to the decision-making. This activation function effectively increases the classification boundary and improves the robustness. In addition, a coarse-grained modal matching loss is added to the model optimization objective to assist the models in learning the cross-modal association features. Experimental results on synthetic datasets show that our proposed method outperforms traditional single-modal detection methods in general, and has advantages over baseline models in terms of accuracy and reliability.

Vibration-Based Discriminant Analysis for Pipeline Leaks Detection

Pipelines are useful for transporting liquids from one place to another. The main problem that often occurs in pipelines is leakage which results in production and financial losses. The importance of detecting pipeline leaks makes the industries look for effective detection methods to avoid bigger losses. Several previous studies have proven that the vibration-based method is successful in detecting leaks in pipelines. However, the vibration-based method used in the previous study is relatively complicated and requires specialists to interpret the results. This study proposes a machine learning-based detection method that can classify pipe conditions directly without the help of a specialist. The proposed method is vibration-based discriminant analysis; a machine learning algorithm that recognizes pipeline conditions from their vibration pattern instead of spectrum. The proposed method was tested on a test rig consisting of a closed-loop pipeline equipped with a leak-pipe test segment. The vibration signal is taken using an accelerometer placed on the leak-pipe test segment. Time domain vibration data is extracted using several statistical parameters which aims to reveal information related to pipe conditions. The vibration data collected is divided into two groups, namely training-data and testing-data. The discriminant analysis model is trained to recognize the vibration pattern of the pipeline using training-data and then tested using testing-data. There are four leak sizes introduced in this study, small, medium, and large. Meanwhile, normal condition (no leaks) is used as benchmarking. The study shows that the proposed method is effective in classifying four pipe conditions with the accuracy up to 95%.

Revealing a New Vulnerability of Distributed State Estimation: A Data Integrity Attack and an Unsupervised Detection Algorithm

This article proposes a distributed false data injection attack (FDIA) by attacking to the boundary buses in an interconnected power system. The proposed attack utilizes the measurements corresponding to a set of boundary buses in each neighboring areas to inject arbitrary errors to the estimated states of those buses. It is demonstrated that the attack not only gets through the robust distributed estimators but also bypasses the convergence-based detection methods. Furthermore, in an illustrative example, the differences in the attack with the conventional FDIA are briefly explained. Then, finding the optimal attack vector to minimize the maximum difference between the per area errors by considering the attacker’s limitations is formulated as a mixed-integer second-order cone programming problem. Finally, an unsupervised machine learning-based detection method is proposed utilizing a kernel density estimation technique along with statistical measures. This follows an outlier detection to filter out attacks. To show the performance of the detector, the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$n-1$</tex-math></inline-formula> contingency, which changes the probability distribution of data is analyzed. The proposed attack and detector are tested on various IEEE systems such as IEEE 14-bus and IEEE 118-bus test systems and the results are discussed.

Survey on Astroturfing Detection and Analysis from an Information Technology Perspective

With the development of the Internet, user comments produced an unprecedented impact on information acquisition, goods purchase, and other aspects. For example, the user comments can quickly render a topic the focus of discussion in social networks. It can promote the sales of goods in e-commerce, and it influences the ratings of books, movies, or albums. Among these network applications and services, “astroturfing,” a kind of online suspicious behavior, can generate abnormal, damaging, and even illegal behaviors in cyberspace that mislead public perception and bring a bad effect on Internet users and society. Hence, the manner of detecting and combating astroturfing behavior has become highly urgent, attracting interest from researchers both from information technology and sociology. In the current paper, we restudy it mainly from the perspective of information technology, summarize the latest research findings of astroturfing detection, analyze the astroturfing feature, classify the machine learning-based detection methods and evaluation criteria, and introduce the main applications. Different from the previous surveys, we also discuss the new future directions of astroturfing detection, such as cross-domain astroturfing detection and user privacy protection.

Detecting Insider Threat from Behavioral Logs Based on Ensemble and Self-Supervised Learning

Recent studies have highlighted that insider threats are more destructive than external network threats. Despite many research studies on this, the spatial heterogeneity and sample imbalance of input features still limit the effectiveness of existing machine learning-based detection methods. To solve this problem, we proposed a supervised insider threat detection method based on ensemble learning and self-supervised learning. Moreover, we propose an entity representation method based on TF-IDF to improve the detection effect. Experimental results show that the proposed method can effectively detect malicious sessions in CERT4.2 and CERT6.2 datasets, where the AUCs are 99.2% and 95.3% in the best case.

Machine Learning-Based Detection Method for Wafer Test Induced Defects

Wafer test is carried out after integrated circuits (IC) fabrication to screen out bad dies. In addition, the results can be used to identify problems in the fabrication process and improve manufacturing yield. However, the wafer test itself may induce defects to otherwise good dies. Test-induced defects not only hurt overall manufacturing yield but also create problems for yield learning, so the source problems in testing should be identified quickly. In the wafer acceptance test process, dies are probed in a predetermined order, so test-induced defects, also known as site-dependent faults, exhibit specific patterns that can be effectively captured in test paths. In this paper, we analyze characteristics of test-induced defect patterns and define features that can be used by machine learning algorithms for the automatic detection of test-induced defects. Therefore, defective dies caused by the wafer test can be retested for yield improvement. Test data from six real products are used to validate the proposed method. Several machine learning algorithms have been applied, and experimental results show that our method is effective to distinguish between test-induced and fabrication-induced defects. On average, the prediction accuracy is higher than 97%.

Detecting damaged buildings from satellite imagery

Especially in recent years, studies to determine the effects of natural disasters from satellite images have been very popular. The destruction caused by the disaster and the early detection of the affected structures are of great importance for the establishment of the precautionary measures and the right action plan. However, studies in this area are mostly made observationally and as a result, desired results cannot be achieved. On the other hand, the introduction of machine learning-based detection methods is very promising. In this study, a damaged building detection method based on convolutional neural networks (CNN) is proposed. Unlike similar studies, the hyperparameters of the CNN are optimized using Bayesian optimization algorithm to obtain more accurate and reliable detection results. The testing and validation results performed with a large number of images reveal the robustness of the proposed method. In addition, the performance evaluation measures obtained from the balanced and unbalanced testing datasets solidified the success of the optimized CNN model.

Lexical features based malicious URL detection using machine learning techniques

Most sophisticated cyber-attack technique used by the cyber criminals is creating and spreading malicious domain names or malicious URLs through email, messages, popups etc. Malicious URL are the web pages targeted towards the internet user to spread the malware, virus, warms etc once the user visited. Main intension of the attack is to steal the victim information, user credentials or install the malware in the victim’s system. So, it is necessary to adapt the system which should detect the malicious URLs and prevent from the attack. Researchers suggest numerous methods but machine learning based detection method performs better then methods. This paper presents the light weighted method which includes only lexical features of the URL. The result shows the Random Forest classifier performs better than the other classifiers in terms of accuracy.

Cost-Based Heterogeneous Learning Framework for Real-Time Spam Detection in Social Networks With Expert Decisions

With the widespread use of social networks, spam messages against them have become a major issue. Spam detection methods can be broadly divided into expert-based and machine learning-based detection methods. When experts participate in spam detection, the detection accuracy is fairly high. However, this method is highly time-consuming and expensive. Conversely, methods using machine learning have the advantage of automation, but their accuracy is relatively low. This paper proposes a spam-detection framework that combines and fully exploits the advantages of both methods. To reduce the workload of the experts, all messages are first analyzed via a primary machine learning filter, and those that are determined to be normal messages are allowed through, whereas suspicious messages are flagged. The flagged messages are subsequently analyzed by an expert to enhance the overall system accuracy. In the filtering process, cost-based machine learning is used to prevent the fatal error of misidentifying a spam message as a normal message. In addition, to obviate the continuously evolving spam trends, a module that periodically updates the expert-diagnosis results on the training dataset is incorporated into the framework. The results of experiments conducted, on an imbalanced dataset of spam tweets and normal tweets in a ratio similar to the actual situation in real life, indicate that the proposed framework has a spam-detection rate of almost 92.8%, which is higher than that of the conventional machine learning technique. Furthermore, the proposed framework delivered stable high performance even in an environment where social network messages changed continuously, unlike the conventional technique, which exhibited large performance deviations.

A Review of Research Work on Network-Based SCADA Intrusion Detection Systems

Specific intrusion detection systems (IDSs) are needed to secure modern supervisory control and data acquisition (SCADA) systems due to their architecture, stringent real-time requirements, network traffic features and specific application layer protocols. This article aims to contribute to assess the state-of-the-art, identify the open issues and provide an insight for future study areas. To achieve these objectives, we start from the factors that impact the design of dedicated intrusion detection systems in SCADA networks and focus on network-based IDS solutions. We propose a structured evaluation methodology that encompasses detection techniques, protected protocols, implementation tools, test environments and IDS performance. Special attention is focused on assessing implementation maturity as well as the applicability of each surveyed solution in the Future Internet environment. Based on that, we provide a brief description and evaluation of 26 selected research papers, published in the period 2015-2019. Results of our analysis indicate considerable progress regarding the development of machine learning-based detection methods, implementation platforms, and to some extent, sophisticated testbeds. We also identify research gaps and conclude the analysis with a list of the most important directions for further research.

Network Anomaly Detection by Using a Time-Decay Closed Frequent Pattern

Anomaly detection of network traffic flows is a non-trivial problem in the field of network security due to the complexity of network traffic. However, most machine learning-based detection methods focus on network anomaly detection but ignore the user anomaly behavior detection. In real scenarios, the anomaly network behavior may harm the user interests. In this paper, we propose an anomaly detection model based on time-decay closed frequent patterns to address this problem. The model mines closed frequent patterns from the network traffic of each user and uses a time-decay factor to distinguish the weight of current and historical network traffic. Because of the dynamic nature of user network behavior, a detection model update strategy is provided in the anomaly detection framework. Additionally, the closed frequent patterns can provide interpretable explanations for anomalies. Experimental results show that the proposed method can detect user behavior anomaly, and the network anomaly detection performance achieved by the proposed method is similar to the state-of-the-art methods and significantly better than the baseline methods.

Danger-Pose Detection System Using Commodity Wi-Fi for Bathroom Monitoring

A bathroom has higher probability of accidents than other rooms due to a slippery floor and temperature change. Because of high privacy and humidity, we face difficulties in monitoring inside a bathroom using traditional healthcare methods based on cameras and wearable sensors. In this paper, we present a danger-pose detection system using commodity Wi-Fi devices, which can be applied to bathroom monitoring, preserving privacy. A machine learning-based detection method usually requires data collected in target situations, which is difficult in detection-of-danger situations. We therefore employ a machine learning-based anomaly-detection method that requires a small amount of data in anomaly conditions, minimizing the required training data collected in dangerous conditions. We first derive the amplitude and phase shift from Wi-Fi channel state information (CSI) to extract low-frequency components that are related to human activities. We then separately extract static and dynamic features from the CSI changes in time. Finally, the static and dynamic features are fed into a one-class support vector machine (SVM), which is used as an anomaly-detection method, to classify whether a user is not in bathtub, bathing safely, or in dangerous conditions. We conducted experimental evaluations and demonstrated that our danger-pose detection system achieved a high detection performance in a non-line-of-sight (NLOS) scenario.

실내 환경에서 검출 속도 개선을 위한 2D 영상에서의 사람 크기 예측

사람 검출의 성능은 카메라의 위치 및 각도 등에 큰 영향을 받는다. 이로 인해 획득한 2D 영상에서 사람은 위치에 따라 각기 다른 크기를 갖는 형태로 나타난다. 이렇게 다양한 크기를 갖는 사람들을 모두 검출하는 것은 실시간 시스템의 구현을 어렵게 만드는 요인이 된다. 그러나 만일 영상의 특정 위치의 사람의 크기를 예측할 수 있다면, 해당 위치의 사람 검출을 위한 연산량이 크게 감소될 수 있을 것이다. 본 논문에서는 실내 공간의 구조를 깊이맵으로 구성하고, 실내 공간에 존재하는 사람의 영상을 3D 공간에 재구성함으로써 크기를 예측하는 기법을 제안한다. 3D 공간에서는 어느 위치에서든지 사람의 크기가 일관되므로 이를 2D 영상으로 투영하게 되면 2D 영상의 좌표에 따른 정확한 사람의 크기를 추정할 수 있다. 실험 결과로부터 제안 방법이 효과적으로 사람의 크기를 예측할 수 있고, 기존이 기계학습 기반 사람 검출 방법들의 처리속도가 감소됨을 증명하였다. The performance of human detection system is affected by camera location and view angle. In 2D image acquired from such camera settings, humans are displayed in different sizes. Detecting all the humans with diverse sizes poses a difficulty in realizing a real-time system. However, if the size of a human in an image can be predicted, the processing time of human detection would be greatly reduced. In this paper, we propose a method that estimates human size by constructing an indoor scene in 3D space. Since the human has constant size everywhere in 3D space, it is possible to estimate accurate human size in 2D image by projecting 3D human into the image space. Experimental results validate that a human size can be predicted from the proposed method and that machine-learning based detection methods can yield the reduction of the processing time.

A machine learning-based method to detect fluorescent spots and an accelerated, parallel implementation of this method

Under most microscopes, the fluorescent signals emitted from biological structures, such as vesicles, as well as the signals from single molecules will appear as small puncta, which contribute to a Gaussian-like distribution. Accurate segmentation of these spots will fundamentally affect our interpretation of a specific biological progress. Because of the complicated backgrounds in images, many algorithms fail to identify all of the interesting signals; the tremendous amount of time required for algorithms to process large datasets can also decrease their utility. Here, we introduce an excellent robust detection method based on the machine learning algorithm AdaBoost, which outperforms threshold-based segmentation, wavelets, and FDA under most situations. We also provide a GPU/multi-core CPU implementation of this algorithm; this implementation accelerates the algorithm approximately 10- and 7-fold acceleration compared with a single CPU implementation. The great reduction of time should make this method a promising candidate in the processing of large datasets. Furthermore, we demonstrate the use of our algorithm on true fluorescent micrographs, and the results show that machine learning-based detection methods outperform the four other previously reported methods.

Graph-based malware detection using dynamic analysis

We introduce a novel malware detection algorithm based on the analysis of graphs constructed from dynamically collected instruction traces of the target executable. These graphs represent Markov chains, where the vertices are the instructions and the transition probabilities are estimated by the data contained in the trace. We use a combination of graph kernels to create a similarity matrix between the instruction trace graphs. The resulting graph kernel measures similarity between graphs on both local and global levels. Finally, the similarity matrix is sent to a support vector machine to perform classification. Our method is particularly appealing because we do not base our classifications on the raw n-gram data, but rather use our data representation to perform classification in graph space. We demonstrate the performance of our algorithm on two classification problems: benign software versus malware, and the Netbull virus with different packers versus other classes of viruses. Our results show a statistically significant improvement over signature-based and other machine learning-based detection methods.