Vietnam Journal of Computer Science

Facial emotion recognition (FER) is the technology or process of identifying and interpreting human emotions based on the analysis of facial expressions. It involves using computer algorithms and machine learning techniques to detect and classify emotional states from images or videos of human faces. Further, FER plays a vital role in recognizing and understanding human emotions to better interpret someone’s feelings, intentions, and attitudes. In the present time, it is widely used in various fields such as healthcare, human–computer interaction, law enforcement, security, and beyond. FER has a wide range of practical applications across various industries including Emotion Monitoring, Adaptive Learning, and Virtual Assistants. This paper presents a comparative analysis of FER algorithms, focusing on deep learning approaches. The performance of different datasets, including FER2013, JAFFE, AffectNet, and Cohn–Kanade, is evaluated using convolutional neural networks (CNNs), deep face, attentional convolutional networks (ACNs), and deep belief networks (DBNs). Among the tested algorithms, DBNs outperformed other algorithms, reaching the highest accuracy of 98.82%. These results emphasize the effectiveness of deep learning techniques, particularly DBNs, in FER. Additionally, outlining the advantages and disadvantages of current research on facial emotion identification might direct future research efforts in the direction of the most profitable directions.

Nowadays, YOLOv5 is one of the most popular object detection network architectures used in real-time and industrial systems. Traffic management and regulation are typical applications. To take advantage of the YOLOv5 network and develop a parking management tool, this paper proposes a car detection network based on redesigning the YOLOv5 network architecture. This research focuses on network parameter optimization using lightweight modules from EfficientNet and PP-LCNet architectures. On the other hand, this work also presents an aerial view dataset for car detection tasks in the parking, named the AVPL. The proposed network is trained and evaluated on two benchmark datasets which are the Car Parking Lot Dataset and the Pontifical Catholic University of Parana+ Dataset and one proposed dataset. The experiments are reported on mAP@0.5 and mAP@0.5:0.95 measurement units. As a result, this network achieves the best performances at 95.8%, 97.4%, and 97.0% of mAP@0.5 on the Car Parking Lot Dataset, the Pontifical Catholic University of Parana+ Dataset, and the proposed AVPL dataset, respectively. A set of demonstration videos and the proposed dataset are available here: https://bit.ly/3YUoSwi.

The categorization of texture images requires the identification and extraction of meaningful keypoints, a crucial step in ensuring the precise representation of textured images. The literature has introduced numerous descriptors in order to detect and capture both local and global texture characteristics. These descriptors vary in their effectiveness depending on the specific application. However, it is generally accepted that they complement each other by compensating for their strengths and weaknesses. Therefore, many approaches focused on combining different types of image descriptors generating robust features that are invariant to image transformations. These solutions mostly focused on one way to deal with information and faced more problems as they need human intervention and a large set of data. The acknowledged benefits of combining multiple types of descriptors are accompanied by challenges specially. These challenges arise due to differences in properties, such as locality and sparsity, as well as the heterogeneity exhibited by generated features. To address this issue comprehensively, we present an approach that benefits from genetic programming techniques to generate and combine two distinct texture classifiers. It incorporates histograms of oriented gradients and local binary patterns descriptors, which capture different textures. To effectively fuse the results of both classifiers, the proposed approach employs a late fusion process along with data augmentation methods while using a limited amount of training data. To evaluate the performance of the proposed approach in texture image classification tasks, we have conducted extensive experiments on six challenging datasets encompassing various variations. We have also investigated its performance in a cross dataset problem where the model has been trained on instances of a dataset before being tested on samples of another dataset. The obtained results clearly demonstrate that the proposed approach surpasses other relevant low-level approaches, as well as existing GP-based and CNN methods specifically designed for describing and classifying textures. Thanks to its ability to simultaneously leverage multiple descriptors, the suggested solution shows a high potential for real-world applications, particularly in handling various image changes with robustness.

Subsequence join in time series is to search for couples of similar subsequences from multiple time series. The task is useful in data mining on time series; nevertheless, it is extremely difficult because of its enormous computational cost. The task should use normalized time series during the search execution and be performed under an efficient distance measure to obtain accurate resulting couples. The task is more challenging when it works in a streaming environment in which time-series data might be collected very quickly. To address this problem, we propose an efficient method of subsequence join in streaming time series under Dynamic Time Warping (DTW), supporting z-score normalization. The proposed method utilizes a technique of subsequence extraction based on major extrema of streaming time series to search for couples of similar subsequences from coevolving time series. This method can identify couples of similar subsequences of the same length or different lengths. The experimental results show that the proposed method has high performance and can bring out interesting couples of similar subsequences. In addition, this method acts as an approximation algorithm suitable for a streaming scenario where users often expect fast responses from the task of subsequence join over time-series streams of high rates.

With the rapid evolution of mobile malware, especially Android malware, machine learning (ML)-based Android malware detection systems have drawn massive attention. Although ML algorithms have recently led to many vital breakthroughs in malware detection, they are still particularly vulnerable to adversarial example (AE) attacks. By applying small random perturbations (e.g. simply modifying different kinds of features from the application’s manifest file), an AE attack can cause the misclassification of legitimate applications. This paper proposes AAGAN, an automated Android malware generation system based on Generative Adversarial Networks (GAN) that can successfully deceive current ML detectors. Our experiment results indicate that AEs generated by our system can flip the prediction of the state-of-the-art detection algorithms in 99% of cases using a real-world dataset. To defend against AE attacks, we improve the robustness of our detection system by alternatively retraining with these newly generated AEs. Surprisingly, after retraining five times, AAGAN can achieve an 89% success rate in bypassing our malware detection system.

The IOTA Tangle, a Directed Acyclic Graph (DAG)-based distributed ledger, is popular for its scalability and suitability for IoT applications, offering fee-less transactions. A critical component of IOTA’s architecture is the Cumulative Weight Calculation (CWC), essential for its tip selection mechanism. This paper introduces an optimization of the IOTA Reference Implementation (IRI) CWC process originally implemented using Breadth-First Search (BFS) by employing Depth-First Search (DFS) and Iterative Deepening Search (IDS) algorithms. We present a comparative analysis of these methods, demonstrating that DFS and IDS provide significant improvements in computational efficiency, particularly beneficial for IoT devices with limited processing capabilities. Our findings are substantiated through a series of experiments on a Tangle snapshot, highlighting the enhanced performance and reduced resource utilization of the proposed methods. This study contributes to the ongoing development of DAG-based distributed ledgers, offering insights into more efficient algorithmic solutions for large-scale, decentralized networks.

Today, every aspect of society’s activities is regulated by laws from the national and local levels. The increasing number of legal documents interwoven with each other also leads to difficulties in searching and applying in practice. In addition, later legal documents may change or invalidate previous legal documents. The construction of knowledge maps that involve one or a group of legal documents is an effective approach to represent actual knowledge domains. A legal knowledge graph constructed from laws and legal documents can enable a number of applications, such as question answering, document similarity, legal search, legal information retrieval, and legal argument claims. In this paper, we describe an ontology-based knowledge map approach for building a Vietnamese legal linked data from the source of about 325,000 legal documents. This study includes an integrated ontology to represent the legal knowledge from legal documents. This model integrates the ontology of relational knowledge and the graph of entities in the form of a concept graph. It can express the semantics of the content of a given legal document. In addition, this study also addresses potential legal applications, which can use the built linked data as a foundation resource of legal. We also highlight open challenges in the realization of knowledge graphs in a technical legal system that enables this approach at scale.