The rapid expansion of Internet of Things (IoT) systems and networks has led to increased challenges regarding security and system reliability. Anomaly detection has become a critical task for identifying system flaws, cyberattacks, and failures in IoT environments. This study proposes a hybrid deep learning (DL) approach combining Autoencoders (AE) and Long Short-Term Memory (LSTM) networks to detect anomalies in real-time within IoT networks. In this model, normal data trends were learned in an unsupervised manner using an AE, while temporal dependencies in time-series data were captured through the use of an LSTM network. Experiments conducted on publicly available IoT datasets, namely the Kaggle IoT Network Traffic Dataset and the Numenta Anomaly Benchmark (NAB) dataset, demonstrate that the proposed hybrid model outperforms conventional machine learning (ML) algorithms, such as Support Vector Machine (SVM) and Random Forest (RF), in terms of accuracy, precision, recall, and F1-score. The hybrid model achieved a recall of 96.2%, a precision of 95.8%, and an accuracy of 97.5%, with negligible false negatives and false positives. Furthermore, the model is capable of handling real-time data with a latency of just 75 milliseconds, making it suitable for large-scale IoT applications. The performance evaluation, which utilized a diverse set of anomaly scenarios, highlighted the robustness and scalability of the proposed model. The Kaggle IoT Network Traffic Dataset, consisting of approximately 630,000 records across six months and 115 features, along with the NAB dataset, which includes around 365,000 sensor readings and 55 features, provided comprehensive data for evaluating the model’s effectiveness in real-world conditions. These findings suggest that the hybrid DL framework offers a robust, scalable, and efficient solution for anomaly detection in IoT networks, contributing to enhanced system security and dependability.

Accurate and efficient detection of small-scale targets on dynamic water surfaces remains a critical challenge in the deployment of unmanned surface vehicles (USVs) for maritime applications. Complex background interference—such as wave motion, sunlight reflections, and low contrast—often leads to missed or false detections, particularly when using conventional convolutional neural networks. To address these issues, this study introduces LMS-YOLO, a lightweight detection framework built upon the YOLOv8n architecture and optimized for real-time marine object recognition. The proposed network integrates three key components: (1) a C2f-SBS module incorporating StarNet-based Star Blocks, which streamlines multi-scale feature extraction while reducing parameter overhead; (2) a Shared Convolutional Lightweight Detection Head (SCLD), designed to enhance detection precision across scales using a unified convolutional strategy; and (3) a Mixed Local Channel Attention (MLCA) module, which reinforces context-aware representation under complex maritime conditions. Evaluated on the WSODD and FloW-Img datasets, LMS-YOLO achieves a 5.5% improvement in precision and a 2.3% gain in mAP@0.5 compared to YOLOv8n, while reducing parameter count and computational cost by 37.18% and 34.57%, respectively. The model operates at 128 FPS on standard hardware, demonstrating its practical viability for embedded deployment in marine perception systems. These results highlight the potential of LMS-YOLO as a deployable solution for high-speed, high-accuracy marine object detection in real-world environments.

When public safety is considered to be of paramount importance, the capacity to detect violent situations through audio monitoring has become increasingly indispensable. This paper proposed a hybrid audio text violence detection system that combines text-based information with frequency-based features to improve accuracy and reliability. The two core models of the system include a frequency-based model, Random Forest (RF) classifier, and a natural language processing (NLP) model called Bidirectional Encoder Representations from Transformers (BERT). RF classifier was trained on Mel-Frequency Cepstral Coefficients (MFCCs) and other spectrum features, whereas BERT identified violent content in transcribed speech. BERT model was improved through task-specific fine-tuning on a curated violence-related text dataset and balanced with class-weighting strategies to address category imbalance. This adaptation enhanced its ability to capture subtle violent language patterns beyond general purpose embeddings. Furthermore, a meta-learner ensemble model using eXtreme Gradient Boosting (XGBoost) classifier model could combine the probability output of the two base models. The ensemble strategy proposed in this research differed from conventionally multimodal fusion techniques, which depend on a single strategy, either NLP or audio. The XGBoost fusion model possessed the qualities derived from both base models to improve classification accuracy and robustness by creating an ideal decision boundary. The proposed system was supported by a Graphical User Interface (GUI) for multiple purposes, such as smart city applications, emergency response, and security monitoring with real-time analysis. The proposed XGBoost ensemble model attained an overall accuracy of over 97.37%, demonstrating the efficacy of integrating machine learning-based decision.