With the rapid growth of online transactions, phishing attacks have become increasingly unpredictable, particularly those involving malicious Uniform Resource Locators (URLs). Currently, few studies have effectively combined manual and semantic features within a unified framework to both fully utilize the structural information and capture the deep context dependence of the character string. To address this gap, this study proposes a deep attention-based approach for phishing URL detection. Our approach first relies on the importance of manual features and the Bidirectional Encoder Representations from Transformers (BERT) model to extract semantic feature vectors. Subsequently, a hybrid deep learning architecture comprising Bidirectional Long Short-Term Memory (BiLSTM) layers, an attention mechanism, and fully connected dense networks is employed to classify URLs as either phishing or legitimate. Experimental results demonstrate that our proposed model achieves a high classification accuracy of 96.77%, while ablation analysis highlights the individual contributions of key components, including BERT embeddings, attention mechanism, LSTM layers, and feature types, to overall model performance. Finally, by training on a Kaggle benchmark dataset and testing on real-world phishing samples, the study confirms the model's strong generalization capability in detecting emerging phishing threats.

Deep learning applications, especially multilayer neural network models, result in network intrusion detection with high accuracy. This study proposes a model that combines a multilayer neural network with Dense Sparse Dense (DSD) multi-stage training to simultaneously improve the criteria related to the performance of intrusion detection systems on a comprehensive dataset UNSW-NB15. We conduct experiments on many neural network models such as Recurrent Neural Network (RNN), Long-Short Term Memory (LSTM), Gated Recurrent Unit (GRU), etc. to evaluate the combined efficiency with each model through many criteria such as accuracy, detection rate, false alarm rate, precision, and F1-Score.

In the context of the COVID-19 pandemic and the risk of similar infectious diseases, monitoring and promoting public health measures like wearing face masks have become crucial in controlling virus transmission. Deep learning-based mask recognition systems play an important role, but their effectiveness depends on the quality and diversity of training datasets. This study proposes the diverse and robust dataset for face mask detection (DRFMD), designed to address limitations of existing datasets and enhance mask recognition models' performance. DRFMD integrates data from sources such as AIZOO, face mask detector by Karan-Malik (KFMD), masked faces (MAFA), MOXA3K, properly wearing masked face detection dataset (PWMFD), and the Zalo AI challenge 2022, comprising 14,727 images with 29,846 instances, divided into training, validation, and testing sets. The dataset's scale and diversity ensure higher accuracy and better generalization for mask recognition models. Experiments with variations of the YOLOv8 model (n, s, m, l, x), an advanced object detection algorithm, on the DRFMD dataset, demonstrate superior performance through metrics like precision, recall, and mAP@50. Additionally, comparisons with previous dataset like FMMD show that models trained on DRFMD maintain strong generalization capabilities and higher performance. This study significantly contributes to improving accuracy of public health monitoring systems, aiding in the prevention of hazards from infectious diseases and air pollution.