JMST's Section on Computer Science and Control Engineering.

A method for bee activities recognition from videos captured at the beehive entrance

Nhung Le — 2024-12-30

Honeybees play an important role in the ecosystem and agricultural economy. To maintain and develop healthy bee colonies, monitoring and recognizing bee activities at the beehive entrance is necessary. In this research, we extend the method in [6] to track and recognize the flight-in and flight-out activities of both pollen-bearing and non-pollen-bearing bees in videos recorded at the beehive entrance. To achieve this goal, a framework consisting of bee detection, bee tracking, and bee activity recognition is proposed. In the first step, to address the imbalance between the number of pollen-bearing and non-pollen-bearing bees, we employed a detection method combining YOLOv5 and the focal loss function. Subsequently, in the tracking step, based on the detection results of the first step, two OC-SORT-based trackers were initialized to determine the trajectories of pollen-bearing and non-pollen-bearing bees. Finally, in the activity recognition step, rules are applied to the tracked trajectories to determine the instantaneous activity states of honey bees and to recognize their overall activities. The experimental results show that the detection step obtained an overall precision of 0.972 whereas the tracking step achieved HOTA values of 77.28%, MOTA of 90.09%, and MOTP of 84.98%.

An efficient decoding of RS-BCH product codes using hybrid root-finding of the polynomial over finite fields

Pham Khac Hoan — 2024-12-30

This paper proposes an efficient decoding method of product codes with component codes being BCH and Reed-Solomon codes characterized by low complexity and latency achieved through parallel computation and the hybrid approach to solving the error locator polynomial over finite fields. The proposed method can be implemented on low-cost hardware platforms, making it suitable for applications in communication systems that highly require reliability and latency.

State feedback sliding mode control SFSMC with pole assignment for single input bilinear systems

Dr Huy Vu Quoc — 2024-12-30

This paper presents a method of synthesizing the state feedback sliding mode controller SFSMC for single-input bilinear systems. The existence of the sliding mode is synthesized by the Lyapunov stability criterion. The sliding surface is designed using the pole assignment method. Here, the novel contribution is that system quality has been tuned by only adjusting one parameter which is completely independent of system dynamics and states. The simulation indicates that the system is stable, and its states converge to zero. The system phase trajectory quickly settles when the parameter is well adjusted.

Coffee leaf rust disease detection using MobileNetV2-based feature extractor, SVM classifier and visualization technique

Le Thi Thu Hong — 2024-12-30

The coffee plant is a vital crop, particularly in Vietnam, and is vulnerable to weather, cultivation methods, and diseases like rust disease. Early detection and treatment of rust disease are essential to ensuring coffee yield and quality. This study introduces a hybrid model for automated rust disease detection from coffee leaf images. The approach employs MobileNetV2 for feature extraction using convolutional neural networks (CNNs) and a Support Vector Machine (SVM) for classification. Experiments also evaluated other lightweight CNNs like MobileNet and NASNetMobile, as well as classifiers like DecisionTree and RandomForest, but MobileNetV2 and SVM delivered optimal results. The model was trained on the publicly available RoCoLe dataset and achieved a rust disease detection accuracy of 97.13%, surpassing standard CNN approaches by 2.39%. Additionally, the study uses Grad-CAM to visualize key areas in coffee leaf images that influence the classification process, offering insights into how the model distinguishes between healthy and diseased leaves. This methodology supports early disease detection and provides tools for understanding the model's decision-making process, contributing to more effective coffee plant disease management.

A style transfer-based augmentation approach for detecting military camouflaged object

Thi Thu Hang Truong — 2024-12-30

Detecting camouflaged objects in military environments is particularly challenging due to the deliberate blending of targets with their surroundings. Despite advancements in deep learning, the limited availability of training data remains a significant obstacle. To address this issue, this paper proposes a novel data augmentation approach that combines multiple style transfer models to transform the style of training images into diverse reference styles. This enriches the training data by simulating various environmental textures and patterns. Structural Similarity (SSIM) is used as an evaluation metric to select style-transferred images that preserve the best structural similarity, enabling detection models to more effectively learn the distinguishing features of camouflaged objects. Extensive experiments with different style transfer methods and SSIM thresholds show that our augmentation approach significantly enhances the accuracy of state-of-the-art detection algorithms. This approach has the potential to improve object detection in military operations, increasing the reliability and precision of automated surveillance systems.

Application of deep neural networks for military symbol recognition from sketch images

Khắc Điệp Nguyễn — 2024-12-30

The purpose of this research is to test the effectiveness of deep neural networks in recognizing sketch images, particularly military symbols. Sketch images are highly abstract and lack typical features of real images, such as color, background, and environmental details, making the use of deep neural networks a significant challenge. To address this issue, we implement a sketch image recognition model based on Convolutional Neural Networks (CNN). The content of the paper includes designing and describing a new CNN model optimized for symbol recognition from sketch images. We have trained this model on a dataset self-constructed by our team. The training results show that the model has high accuracy in recognizing military symbols from sketch images, confirming the potential of deep neural networks in this field.

FedEC: Enhancing model federated averaging via two-sided method

Do Tuan Minh — 2024-12-30

Federated learning is a key method for addressing data privacy and security in distributed AI training. However, non-IID data among local clients often causes issues like client drift, leading to slow and unstable model convergence. Current studies typically use one-sided methods that optimize either the client or server-side with the conventional aggregation method FedAvg. In contrast, we introduce FedEC, a two-sided strategy that combines distinct methods: an elastic aggregation algorithm for optimizing the global model on the server and contrastive learning techniques on the client side to reduce divergence between local and global models. This complementary approach fosters mutual reinforcement between client and server, allowing FedEC to better tackle non-IID data challenges. Results from experiments conducted on some benchmark datasets with different settings show that FedEC offers more efficient training and outperforms previous one-sided algorithms, underscoring the effectiveness of this two-sided approach in federated learning.

GanTextKnockoff: stealing text sentiment analysis model functionality using synthetic data

Cong Pham — 2024-12-30

Today, black-box machine learning models are often subject to extraction attacks that aim to retrieve their internal information. Black-box model extraction attacks are typically conducted by providing input data and, based on observing the output results, constructing a new model that functions equivalently to the original. This process is usually carried out by leveraging available data from public repositories or synthetic data generated by generative models. Most model extraction attack methods using synthetic data have been concentrated in the field of computer vision, with minimal research focused on model extraction in natural language processing. In this paper, we propose a method that utilizes synthetic textual data to construct a new model with high accuracy and similarity to the original black-box sentiment analysis model.

Intent classification for voice-based military information search on digital maps using integrated BiGRU-CNN network and speech recognition technology

Duc Thinh Dang — 2024-12-30

Searching for information is one of the most important functions of software that supports drafting operational documents on digital maps. To enhance usability and meet the demands of modern military operations, it is necessary to automate the information search function using voice commands. A universal voice search tool that supports searches for various types of information requires an initial step of search intent classification. This paper proposes the development of a search intent classification process using an integrated BiGRU-CNN network and automatic speech recognition technology (ASR). The BiGRU-CNN network leverages the advantages of both BiGRU and CNN models to improve the efficiency of classifying text data converted from speech using the Whisper model. The paper compares the proposed method with those that use separate machine learning models combined with feature extraction methods such as TF-IDF, N-gram, and SVD. While the ASR model used in this research still has constraints, experimental results show that the accuracy of search intent classification reaches up to 98.4%. This result is higher than that of compared methods using simpler machine learning models, demonstrating the effectiveness of the proposed method.

Proposal for an information-hiding model in executable files

Tong Minh Duc — 2024-12-30

In this paper, the authors propose a model for hiding information in executable files on a 64-bit Windows environment based on embedding data into the empty spaces between sections in executable files. After embedding, the information does not increase the file size, does not affect the execution of the original file, and avoids false positives from antivirus programs. This hiding method is based on analyzing the file structures, the mechanism for concealing malicious code into the last section of the executable file, and the mechanism for loading files into executable memory. To enhance the security of the hidden information, data should be encrypted before embedding.

A solution for constructing quantum – resistant digital signature schemes

Lưu Hồng Dũng — 2024-12-30

In this article, the authors propose a solution for constructing quantum - resistant digital signature schemes based on a new type of hard problem, which belongs to the group of unsolvable problems. Therefore, the algorithms constructed according to the solution proposed here can be resistant to quantum attacks based on the quantum algorithm proposed by P. Shor. In addition to quantum resistance, the signature schemes proposed here can also be used as pre-quantum digital signature schemes (RSA, DSA, etc.) that are widely used in current practical applications.