PeerJ Computer Science最新文献

Today, a great number of attack opportunities for cybercriminals arise in Android, since it is one of the most used operating systems for many mobile applications. Hence, it is very important to anticipate these situations. To minimize this problem, the analysis of malware search applications is based on machine learning algorithms. Our work uses as a starting point the features proposed by the DREBIN project, which today constitutes a key reference in the literature, being the largest public Android malware dataset with labeled families. The authors only employ the support vector machine to determine whether a sample is malware or not. This work first proposes a new efficient dimensionality reduction of features, as well as the application of several supervised machine learning algorithms for prediction purposes. Predictive models based on Random Forest are found to achieve the most promising results. They can detect an average of 91.72% malware samples, with a very low false positive rate of 0.13%, and using only 5,000 features. This is just over 9% of the total number of features of DREBIN. It achieves an accuracy of 99.52%, a total precision of 96.91%, as well as a macro average F1-score of 96.99%.

Effective malware detection is critical to safeguarding digital ecosystems from evolving cyber threats. However, the scarcity of labeled training data, particularly for cross-family malware detection, poses a significant challenge. This research proposes a novel architecture ConvNet-6 to be used in Siamese Neural Networks for applying Zero-shot learning to address the issue of data scarcity. The proposed model for malware detection uses the ConvNet-6 architecture even with limited training samples. The proposed model is trained with just one labeled sample per sub-family. We conduct extensive experiments on a diverse dataset featuring Android and Portable Executables' malware families. The model achieves high performance in terms of 82% accuracy on the test dataset, demonstrating its ability to generalize and effectively detect previously unseen malware variants. Furthermore, we examine the model's transferability by testing it on a portable executable malware dataset, despite being trained solely on the Android dataset. Encouragingly, the performance remains consistent. The results of our research showcase the potential of deep convolutional neural network (CNN) in Siamese neural networks for the application of zero-shot learning to detect cross-family malware, even when dealing with minimal labeled training data.

Background: The best possible treatment planning and patient care depend on the precise diagnosis of brain diseases made with medical imaging information. Magnetic resonance imaging (MRI) is increasingly used in clinical score prediction and computer-aided brain disease (BD) diagnosis due to its outstanding correlation. Most modern collaborative learning methods require manually created feature representations for MR images. We present an effective iterative method and rigorously show its convergence, as the suggested goal is a non-smooth optimization problem that is challenging to tackle in general. In particular, we extract many image patches surrounding these landmarks by using data to recognize discriminative anatomical characteristics in MR images. Our experimental results, which demonstrated significant increases in key performance metrics with 500 data such as specificity of 94.18%, sensitivity of 93.19%, accuracy of 96.97%, F1-score of 94.18%, RMSE of 22.76%, and execution time of 4.875 ms demonstrated the efficiency of the proposed method, Deep Multi-Task Convolutional Neural Network (DMTCNN).

Methods: In this research present a DMTCNN for combined regression and classification. The proposed DMTCNN model aims to predict both the presence of brain diseases and quantitative disease-related measures like tumor volume or disease severity. Through cooperative learning of several tasks, the model might make greater use of shared information and improve overall performance. For pre-processing system uses an edge detector, which is canny edge detector. The proposed model learns many tasks concurrently, such as categorizing different brain diseases or anomalies, by extracting features from image patches using convolutional neural networks (CNNs). Using common representations across tasks, the multi-task learning (MTL) method enhances model generalization and diagnostic accuracy even in the absence of sufficient labeled data.

Results: One of our unique discoveries is that, using our datasets, we verified that our proposed algorithm, DMTCNN, could appropriately categorize dissimilar brain disorders. Particularly, the proposed DMTCNN model achieves better than state-of-the-art techniques in precisely identifying brain diseases.

Traditional biometric techniques often require direct subject participation, limiting application in various situations. In contrast, gait recognition allows for human identification via computer analysis of walking patterns without subject cooperation. However, occlusion remains a key challenge limiting real-world application. Recent surveys have evaluated advances in gait recognition, but only few have focused specifically on addressing occlusion conditions. In this article, we introduces a taxonomy that systematically classifies real-world occlusion, datasets, and methodologies in the field of occluded gait recognition. By employing this proposed taxonomy as a guide, we conducted an extensive survey encompassing datasets featuring occlusion and explored various methods employed to conquer challenges in occluded gait recognition. Additionally, we provide a list of future research directions, which can serve as a stepping stone for researchers dedicated to advancing the application of gait recognition in real-world scenarios.

Microbial proliferation presents a significant challenge in membrane-based water for injection (WFI) production, particularly in systems with storage and ambient distribution, commonly refered to as cold WFI production. A comprehensive microbial risk assessment of membrane-based WFI systems was performed by employing Fuzzy-Failure Mode and Effects Analysis (Fuzzy-FMEA) to evaluate the potential microbial risks. Failure modes were identified and prioritized based on the Risk Priority Number (RPN), with appropriate preventive measures recommended to control failure modes that could increase the microbial load and mitigate their impact. Key hazards were identified including fouling of ultrafiltration (UF) membranes, insufficient sealing of heat exchangers, leakage in reverse osmosis (RO) membranes, and ineffective vent filters unable to remove airborn microorganism. Based on Fuzzy-FMEA results, suggestions for optimization were proposed to improve microbial control in membrane-based WFI systems in the pharmaceutical industry.

Emotion recognition is a significant research problem in affective computing as it has a lot of potential areas of application. One of the approaches in emotion recognition uses electroencephalogram (EEG) signals to identify the emotion of a person. However, effectively using the global and local features of EEG signals to improve the performance of emotion recognition is still a challenge. In this study, we propose a novel Convolution Interactive Transformer Network for EEG Emotion Recognition, known as CIT-EmotionNet, which efficiently integrates the global and local features of EEG signals. We convert the raw EEG signals into spatial-spectral representations, which serve as the inputs into the model. The model integrates convolutional neural network (CNN) and Transformer within a single framework in a parallel manner. We propose a Convolution Interactive Transformer module, which facilitates the interaction and fusion of local and global features extracted by CNN and Transformer respectively, thereby improving the average accuracy of emotion recognition. The proposed CIT-EmotionNet outperforms state-of-the-art methods, achieving an average recognition accuracy of 98.57% and 92.09% on two publicly available datasets, SEED and SEED-IV, respectively.

Background: Ship detection in video surveillance images holds significant practical value. However, the background in these images is often complex, complicating the achievement of an optimal balance between detection precision and speed.

Method: This study proposes a ship detection method that leverages semantic aggregation in complex backgrounds. Initially, a semantic aggregation module merges deep features, rich in semantic information, with shallow features abundant in location details, extracted via the front-end network. Concurrently, these shallow features are reshaped through the reorg layer to extract richer feature information, and then these reshaped shallow features are integrated with deep features within the feature fusion module, thereby enhancing the capability for feature fusion and improving classification and positioning capability. Subsequently, a multiscale object detection layer is implemented to enhance feature expression and effectively identify ship objects across various scales. Moreover, the distance intersection over union (DIoU) metric is utilized to refine the loss function, enhancing the detection precision for ship objects.

Results: The experimental results on the SeaShips dataset and SeaShips_enlarge dataset demonstrate that the mean average precision@0.5 (mAP@0.5) of this proposed method reaches 89.30% and 89.10%, respectively.

Conclusions: The proposed method surpasses other existing ship detection techniques in terms of detection effect and meets real-time detection requirements, underscoring its engineering relevance.

Globally, students with visual disabilities face significant challenges in accessing and learning mathematics, particularly when solving mathematical equations and expressions. These challenges result from the inherent complexity and abstract nature of mathematical content. Additionally, braille codes are inconsistent across regions, collaborative math platforms are unavailable, and accessible mathematics literature is scarce. Assistive technologies, artificial intelligence, and educational resources have improved accessibility for students with visual disabilities. However, these students still face significant challenges when navigating, exploring, and solving mathematical equations and expressions. These challenges contribute underrepresentation of these students in the science, technology, engineering, and mathematics disciplines. To address these limitations, this study proposes a novel solution to assist students with visual disabilities in learning mathematical expressions interactively with flexible navigation. This study proposes an algorithmic approach for converting input mathematical expressions into content MathML expressions, parsing those expressions into semantic elements, and then providing a structural overview of these expressions. Moreover, interactive keyboard keys were designed to provide flexible navigation through speech feedback, so that users can interact more effectively with expressions. Python libraries were utilized to implement the proposed solution. An empirical evaluation was conducted by 15 instructors and 94 students with visual disabilities and validated by Cronbach's alpha. Results indicate that the proposed solution improved mathematics accessibility and learning. This study lays a foundation for future research on the integration of advanced technologies in special education.

Crowd management refers to the management and control of masses at specific locations. A Hajj gathering is an example. Hajj is the biggest gathering of Muslims worldwide. Over two million Muslims from all over the globe come annually to Makkah, Saudi Arabia. Authorities of Saudi Arabia strive to provide comfortable comprehensive services to pilgrims using the latest modern technologies. Recent studies have focused on camera scenes and live streaming to assess the count and monitor the behavior of the crowd. However, the opinions of the pilgrims and their feelings about their experience of Hajj are not well known, and the data on social media (SM) is limited. This paper provides a combination of Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) algorithms for sentiment analysis of pilgrims using a novel and specialized dataset, namely Catering-Hajj. The model is based on four CNN layers for local feature extraction after the One-Hot Encoder, and one LSTM layer to maintain long-term dependencies. The generated feature maps are passed to the SoftMax layer to classify final outputs. The proposed model is applied to a real case study of issues related to pre-prepared food at Hajj 1442. Started with collecting the dataset, extracting target attitudes, annotating the data correctly, and analyzing the positive, negative, and neutral attitudes of the pilgrims to this event. Our model is compared with a set of Machine Learning (ML) models including Support Vector Machine (SVM), Logistic Regression (LR), Decision Tree (DT), and Random Forest (RF), as well as CNN and LSTM models. The experimental results show that SVM, RF, and LSTM achieve the same rate of roughly 81%. LR and CNN achieve 79%, and DT achieves 71%. The proposed model outperforms other classifiers on our dataset by 92%.

The complex environments and unpredictable states within transportation networks have a significant impact on their operations. To enhance the level of intelligence in transportation networks, we propose a visual scene feature clustering analysis method based on 3D sensors and adaptive fuzzy control to address the various complex environments encountered. Firstly, we construct a feature extraction framework for visual scenes using 3D sensors and employ a series of feature processing operators to repair cracks and noise in the images. Subsequently, we introduce a feature aggregation approach based on an adaptive fuzzy control algorithm to carefully screen the preprocessed features. Finally, by designing a similarity matrix for the transportation network environment, we obtain the recognition results for the current environment and state. Experimental results demonstrate that our method outperforms competitive approaches with a mean average precision (mAP) value of 0.776, serving as a theoretical foundation for visual scene perception in transportation networks and enhancing their level of intelligence.