Detecting anomalous objects in aquaculture ponds is an important but challenging problem due to the lack of labeled data. This study proposes an unsupervised learning method combining Panoptic Segmentation, Energy Distance, and Point of Interest (PoI) to automatically detect anomalous objects from pond images. The system extracts a 12-dimensional feature vector specialized for aquatic environments and calculates an Energy Distance Map measuring feature distribution differences between image regions. A two-step segmentation strategy is applied: segmenting objects with low-to-medium energy, then creating instances for high-energy regions corresponding to potential anomalous objects. Classification is based on dynamic thresholds from the 75th percentile, allowing adaptation to different image conditions. Experiments on three diverse scenarios with varying density and lighting showed promising localization results, demonstrating feasibility for smart aquaculture applications.
The demand for precise and efficient object detection in aerial imagery has surged, driven by applications in agriculture, surveillance, disaster management, and environmental monitoring. However, detecting small objects in drone-captured images remains challenging due to factors like low resolution, occlusion, and varying scales. This research explores a novel approach to small, open vocabulary object detection by combining the OWL-ViT (OpenWorld Vision Transformer) model with the SAHI (Slicing Aided Hyper Inference) technique. OWL-ViT, known for its ability to handle open vocabulary object detection, is leveraged for its robust feature extraction and generalization capabilities across diverse object categories. SAHI is integrated to address the small object detection challenge by slicing high resolution drone images into smaller patches, enabling more focused and detailed inference. In a comprehensive evaluation, our combined method achieves significant improvements in mAP@50 for small-scale object detection, with an average increase of +6.8% on the VisDrone dataset.
This paper presents a novel approach to enhancing object detection efficiency by integrating multi-level features within a transformer architecture. Traditional object detection methods often rely on single-level feature representations, which may limit their ability to accurately detect objects of varying sizes and complexities. By leveraging multi-level feature integration within the transformer framework, our method captures a richer set of spatial and semantic information, leading to more precise and robust object detection.The powerful attention mechanisms of transformers are utilized to effectively combine these features, improving detection accuracy and localization. The proposed approach is evaluated on the PASCAL VOC benchmark dataset, demonstrating superior performance over conventional single-level feature-based methods. Experimental results show that our model achieves an mAP@0.5 of 87% on PASCAL VOC, outperforming recent state-of-the-art methods while maintaining computationally efficient. These findings highlight the potential of multi-level feature integration within transformers in advancing the field of object detection.
This paper presents the design and implementation of a practical news aggregation system for decision making and risk management in the banking sector. The system gathers financial news from multiple sources and mines texts to provide insight into banks’ professional activities such as policies, products, and financial performance. Multiple natural language processing (NLP) modules are integrated to address tasks such as topic classification and sentiment analysis. To identify the most suitable technologies for each component, we systematically evaluated a wide range of NLP techniques, including large language models (LLMs) and domain-specific pre-trained models. A Vietnamese financial corpus of 12,000 annotated samples was constructed to fine-tune models such as PhoBERT, ViT5, and BARTPho. Experimental results show that fine-tuned models significantly outperform general-purpose LLMs (e.g., LLaMA-3.1-8B, Vistral-7B) in both accuracy and computational efficiency. The fine-tuned models achieve a 7.15% accuracy improvement and reduce resource requirements. The study demonstrates a scalable and adaptable framework for building multi-source, text-based intelligent systems in the financial domain.
In this paper, by introducing novel conditional games, we contribute to AI and multi-agent systems by modeling real-world coalition formation with multiple interdependent constraints that simple games cannot achieve. While simple games treat all players as identical units with a single threshold, conditional games with linear constraints create geometric structure in the solution space, transforming NP-hard problems into polynomial-time solvable ones. This enables AI systems to identify minimal winning coalitions in real-time for complex scenarios like resource-constrained influence maximization in influencer marketing, cybersecurity configuration management, and ethical AI governance frameworks. The precise boundary analysis provided by linear constraints allows multi-agent systems to navigate strategic thresholds with mathematical precision, optimizing resource allocation by identifying exactly which agents are critical for crossing success boundaries. This computational and strategic advantage, turning constraint-induced structure into navigable geometric landscapes, enables sophisticated coalition formation, predictive strategic planning, and dynamic adaptation that simple games, with their arbitrary winning coalition distributions and exponential complexity, simply cannot support in practical applications. The geometric-structure-leveraged Constraint Projection Algorithm is presented with real-world simulation applied in green influencer marketing, demonstrating how the polynomial constraint projection algorithm transforms influencer selection from a trial-and-error process into a precise, data-driven optimization.
Early identification of stroke symptoms is essential for enabling timely intervention and improving patient outcomes, particularly in prehospital settings. This study presents a fast, non-invasive multimodal deep learning framework for automatic binary stroke screening based on data collected during the F.A.S.T. assessment. The proposed approach integrates complementary information from facial expressions, speech signals, and upper-body movements to enhance diagnostic robustness. Facial dynamics are represented using landmark based features and modeled with a Transformer architecture to capture temporal dependencies. Speech signals are converted into mel spectrograms and processed using an Audio Spectrogram Transformer, while upper-body pose sequences are analyzed with an MLP-Mixer network to model spatiotemporal motion patterns. The extracted modality specific representations are combined through an attention-based fusion mechanism to effectively learn cross modal interactions. Experiments conducted on a self-collected dataset of 222 videos from 37 subjects demonstrate that the proposed multimodal model consistently outperforms unimodal baselines, achieving 95.83
Early detection of skin cancer significantly improves patient outcomes by allowing for timely intervention. This study introduces the FusionNetX framework, which is a robust multimodal model using both image data and metadata for skin cancer detection, leveraging the ISIC 2024 dataset. Our approach integrates convolutional neural networks (CNNs) and Transformer-based models to extract features from single-lesion images cropped from 3D Total Body Photographs (3D-TBP). These images, resembling close-up smartphone photos, are integrated with metadata analyzed using tree-based classifiers to enhance diagnostic accuracy. To address the extreme class imbalance in the dataset, we employ advanced sampling techniques and use stratified group cross-validation to ensure our model generalizes well across diverse patient groups. Our model demonstrates competitive performance, achieving a partial area under the Receiver Operating Characteristic curve (pAUC) of 0.18380 on cross-validation and securing the top rank with a private score of 0.17295 on the private test set. This top-ranking performance highlights the model’s ability to maintain high true positive rates (TPR ≥ 80%) while outperforming all other teams based on private scores. These results underscore the effectiveness of our multimodal approach, offering a promising solution for improving early skin cancer detection and enhancing patient outcomes.
With the increasing adoption of post-quantum cryptographic schemes such as SPHINCS+ in Narrowband IoT (NB-IoT) systems, a major challenge arises from large packet size overhead - often reaching approximate 34 KB per transmission due to stateless hash-based signature metadata. This paper investigates practical techniques for reducing this overhead while maintaining quantum resilience, including signature aggregation across multiple devices, Merkle tree pruning within SPHINCS+, and selective signing policies based on session frequency. We implement these strategies on an ESP32-WROOM-32 microcontroller interfaced with a SIM7080 NB-IoT module, demonstrating that packet size can be reduced by up to 60% through intelligent session management and optimized scheduling. Our results provide actionable insights for system architects aiming to deploy post-quantum security within strict bandwidth constraints typical of LPWAN environments.
Content-based image retrieval (CBIR) is becoming increasingly important amid the rapid growth of image data. Traditional CBIR approaches, which rely on features such as color, shape, and texture, often face limitations in accuracy. Even when using features extracted from deep learning models, these approaches still fall short of fully meeting user expectations. To enhance retrieval effectiveness, this study introduces an ensemble approach that utilizes feature sets from multiple deep learning models. In our method, retrieved images are determined through an aggregation of recommendations from deep learning models, with each model’s vote assigned a specific weight. This weight is comprehensively evaluated based on the similarity between the recommended image and the query, the model’s reliability, and the distribution of images recommended by each model. To validate the effectiveness of this approach, we conducted experiments using VGG16, ResNet50, EfficientNetB0, DenseNet201, Swin, and Clip, pre-trained on ImageNet for feature extraction. Three model combinations, (1) VGG16, ResNet50, and EfficientNetB0, (2) ResNet50, EfficientNetB0, and DenseNet201, and (3) Swin and Clip, were explored within the proposed ensemble framework on the Oxford-17-Flowers, Caltech-101, CIFAR-10, and ISIC-2018 datasets. The suggested approach routinely outperforms individual models, according to experimental results, providing better retrieval accuracy on most datasets.
The increasing sophistication of cyberattacks, causing global damages estimated at $9.22 trillion in 2024, highlights the critical importance of robust Intrusion Detection Systems (IDS). AI-driven IDS frameworks, such as APELID, demonstrate impressive detection accuracy leveraging novel machine learning. However, these systems remain vulnerable to adversarial machine learning (AML) attacks, which craft deceptive inputs to bypass detection mechanisms. In this paper, we propose APELID+, an enhanced IDS framework integrating adversarial training and feature squeezing techniques to effectively counter AML threats. We systematically evaluate APELID’s vulnerabilities using comprehensive adversarial attack strategies, including both white-box (FGSM, JSMA, PGD, DeepFool, CW) and black-box attacks (ZOO, HSJA). Experimental results on the CSE-CIC-IDS2018 dataset reveal a significant reduction in APELID’s accuracy (from 99.7% to as low as 1.14% under FGSM attacks). The enhanced APELID+ achieves robust performance, maintaining 98.73% accuracy under combined adversarial conditions, surpassing state-of-the-art methods such as Apollon and RAIDS.
Fine-grained classification of pulmonary diseases remains a challenging task due to subtle inter-class variations and overlapping visual patterns in Chest X-ray imaging. In this work, we propose FSGCN (Fine-grained Silicosis Graph-based Classification Network), a novel hybrid architecture that combines convolutional representation learning with relational reasoning via a graph transformer network (GTN). Specifically, image features extracted from a deep encoder are treated as nodesin a fully connected graph, where edge relationships are dynamically learned to capture semantic correlations among instances within a batch. This graph-based interaction enables the model to better distinguish silicosis from other visually similar pulmonary conditions, such as viral and bacterial pneumonia. We evaluate the proposed approach on SVBCX, a curated multiclass chest X-ray dataset collected and annotated by our research team, comprising four categories: Normal, Silicosis, Viral, andBacterial. The experimental results demonstrate the superiority of our method over existing baselines.On average, FSGCN achieves an absolute improvement of +1.51% in accuracy and +1.85% in F1-score compared to state-of-the-art baseline models. In addition, qualitative visualizations further confirm the effectiveness of the proposed model in enhancing class discriminability. These results highlight the importance of modeling inter-instance relationships for fine-grained disease classification in medical imaging.
Non-orthogonal multiple access (NOMA) enables simultaneous transmissions of a huge quantity of wireless users on the same system resources (time, frequency, space), significantly improving spectral efficiency. Also, harvesting radio frequency energy for wireless communication not only saves energy resource but also makes communication greener. However, an energy harvesting-capable eavesdropper can be one of NOMA receivers, threatening security of NOMA communication. This paper secures NOMA communication eavesdropped by the energy harvesting-capable receiver by utilizing a jammer. The secrecy performance of the proposed system model is analyzed thoroughly and corroborated by computer simulations. Moreover, its secrecy performance is demonstrated to be superior to three different reference models. Furthermore, we secure the proposed system model at best by properly configuring critical system parameters.
Human Action Recognition (HAR) has emerged as an active research domain in recent years with wide-ranging applications in healthcare monitoring, smart home systems, and hu- man–robot interaction. This paper introduces a method, namely Mamba-MHAR (Mamba based Multimodal Human Action Recognition), a lightweight multimodal architecture aimed at improv- ing HAR performance by effectively integrating data from inertial sensors and egocentric videos. Mamba-MHAR consists of double Mamba-based branches, one for visual feature extraction - VideoMamba, and the other for motion feature extraction - MAMC. Both branches are built upon recently introduced Selective State Space Models (SSMs) to optimize the computational cost, and they are lately fused for final human activity classification. Mamba-MHAR achieves significant efficiency gains in terms of GPU usage, making it highly suitable for real-time deployment on edge and mobile devices. Extensive experiments were conducted on two challenging multimodal datasets UESTC-MMEA-CL and MuWiGes, which contain synchronized IMU and video data recorded in natural settings. The proposed Mamba-MHAR achieves 98.00% accuracy on UESTC-MMEA-CL and 98.58% on MuWiGes, surpassing state-of-the-art baselines. These results demonstrate that a simple yet efficient fusion of multimodal lightweight Mamba-based models provides a promising solution for scalable and low-power applications in pervasive computing environments.
The breakthroughs in natural language processing (NLP) are not only a crucial step in technological evolution but also deliver significant benefits across various fields demanding high intelligence and precision. One of the notable NLP applications is in the analysis and processing of legal texts. Capitalizing on this trend, the 10th Workshop on Vietnamese Language and Speech Processing (VLSP) 2023 hosted a new challenge: Legal textual entailment recognition (RTE). The task involves determining whether a given statement is logically entailed by the relevant legal passage. Our proposed method leverages a novel layer based on Support Vector Machine (SVM) kernel formulations, effectively capturing nuanced relationships in the input data. Additionally, it capitalizes on the advantages of the natural language inference (NLI) datasets which are very close to textual entailment recognition (RTE) for enhancing performance and generalization. Our approach not only yielded accurate results but also demonstrated efficiency in the use of data resources, helping our A3N1 team achieve notable accuracy, with a score of 0.7194 on the test set, and ranking third on the leaderboard.
The Visual Question Answering (VQA) problem represents the fusion of natural language understanding (NLU) and computer vision, aiming to comprehend both textual queries and visual content. Recently, researchers have been focusing on the reading comprehension abilities of VQA models, specifically on their capacity to utilize information from scene texts to gather additional context for answering the posed questions. In this study, we present a fundamental approach for integrating diverse information contained in both images and questions. By leveraging a Transformer model, the proposed solution effectively addresses the VQA problem. Our approach, achieving the second position in the VLSP 2023 challenge on Visual Reading Comprehension for Vietnamese, highlights the effectiveness of our proposed method. This study contributes to the ongoing discourse on refining VQA models and emphasizes the potential for further advancements in this domain. The codes are available in the GitHub repository, i.e., https://github.com /truong-xuan-linh/FSO-implement.
Person re-identification (ReID) plays a crucial role in computer vision-based surveillance systems, enabling the accurate identification of individuals across multiple camera views. Traditional convolutional neural network (CNN)-based approaches, such as those utilizing ResNet-50, struggle to capture long-range dependencies and contextual relationships, limiting their effectiveness in diverse real-world scenarios. To overcome these challenges, recent advancements have explored Vision Transformer (ViT)-based architectures, leveraging self-attention mechanisms for enhanced feature representation. In this research, we introduce a ViT-based framework, namely ViTC-UReID, for unsupervised person ReID by incorporating a camera-aware proxy learning mechanism to improve feature consistency across different camera viewpoints. Moreover, ViTC-UReID also uses clustering algorithms to generate pseudo labels for samples in training datasets. Our approach significantly enhances cross-camera adaptation, reducing domain shift effects while maintaining strong feature discrimination. We evaluate our method on three widely used benchmarks Market-1501, MSMT17, and CUHK03, demonstrating its superior performance compared to existing state-of-the-art unsupervised methods, particularly those utilizing camera identity cues. Furthermore, our model achieves competitive accuracy with fully supervised methods, highlighting the effectiveness of transformer-based representations in complex person ReID scenarios. Our findings reinforce the growing potential of unsupervised person ReID methods and demonstrate that ViT architectures combined with camera-aware learning can drive substantial improvements in person ReID.
In the face of the explosive growth in data volume, current technologies have encountered many difficulties in both storage and knowledge discovery processes. Moreover, the quality of data has also deteriorated due to excessive noisy information, which reduces the effectiveness of machine learning models. Therefore, many solutions have been proposed, in which attribute reduction has emerged as an important research direction. Currently, research on attribute reduction has become very active and is primarily focused on the processing of decision tables. In this area, research on attribute reduction based on rough set theory and its extensions is considered a promising direction, which has been yielding many impressive results. To gain a clearer understanding of the attribute reduction research direction, this study will provide an overview of the methods of attribute reduction from their inception to the methods proposed in recent times.
Learning to Optimize (L2O) is an emerging research area in machine learning, focusing on designing and training optimization algorithms that can learn to improve their own performance through experience. Each inference solves a data-driven optimization problem. L2O models are designed to be easy to deploy, incorporate prior knowledge and ensure correctness, such as satisfaction of constraints. This paper applies L2O with the combination of certificates, achieving a higher level of explainability for AI decisions than previous Explainable AI (XAI) methods on two low-dose CT image reconstruction datasets, LoDoPab and Ellipses. The paper also introduces a method to reduce the number of parameters and training time of the model while maintaining the same performance and ensuring the constraints conditions.
Generative Adversarial Networks (GANs) have recently become an interesting subject for researchers due to their diverse applications across various fields. Initially focused on imagerelated tasks, they then have been used to generate new synthetic data for applications across many areas of machine learning research. In malware analysis, GANs have rapidly expanded and are used to generate adversarial data for enhancing the effectiveness of malware detection systems. This paper reviews the application of GANs in malware analysis to generate adversarial examples, to modify semantic information within data, to augment datasets for rare classes, and to support representation learning. The paper provides an extensive overview that serves both as a primer for cybersecurity specialists and a resource for machine learning researchers. The paper outlines the fundamentals of GANs, their operational mechanism, the current types of GANs, the challenges faced, and future directions for exploration in malware analysis.
Continual learning models encounter a significant challenge, namely catastrophic forgetting. After training on a new task they forget previously learned knowledge, causing a significant drop in models’ performance on previous tasks. Replay-based methods, the most efficient approach for addressing catastrophic forgetting, select a small number of data (the coreset) from each learned task to store into an episodic memory buffer for rehearsal during subsequent stages. This selection significantly impacts the models’ ability to perform well on new tasks while maintaining performance on previous tasks. In this paper, we propose a two-phase method to address catastrophic forgetting in continual learning. The first phase utilizes the second-order influence function to select an effective coreset from previously learned tasks. Even with this effective selection, there is still a problem that the hidden feature space is unexpectedly transformed across each task, causing the model to forget optimal hidden representations on previously learned tasks. To address this issue, the second phase employs an energy-based latent aligner (ELI) to re-align the hidden feature representations of tasks towards the optimal region. Extensive experiments on three continual learning benchmark datasets, i.e. CIFAR-10, CIFAR-100, and Split miniImageNet, demonstrate that our proposed method outperforms several existing state-of-the-art continual learning models.
