Intrusion detection in self-regulating manufacturing requires light yet precise modes to do real-time threat prevention. The performance of TinyML, 1D-CNN, GRU, and LSTM models is tested based on different metrics in this research. TinyML works better than all models with 97% accuracy and log-loss as low as 1.2, which is an indication of confident predictions. One of the significant aspects of TinyML is improved accuracy, and decreased sensitivity, with increased false negative and false positive rates also making it more useful in resource-constrained areas. For natural language generation tasks, Mixtral 8x7B 32768 shows the highest BLEU score of 0.60 and METEOR score of 0.86, reflecting its high similarity with reference outputs. It also scores the highest in ROUGE-1 (0.77) and ROUGE-2 (0.75), providing high-quality phrase-level recall. These results confirm that TinyML is the most effective for intrusion detection, while Mixtral 8x7B 32768 excels in text generation tasks.

This paper introduces a secure federated learning (FL) system aimed at detecting anomalies in the communication of maritime autonomous surface ships (MASS). The model effectively distinguishes between normal and anomalous nano-traffic while safeguarding data privacy among distributed clients. To bolster security, Quantum Key Distribution (QKD) protocols, such as BB84 and E91, are utilized for encrypted key exchange during ship-to-ship and ship-to-ground interactions. The suggested framework demonstrates consistent performance enhancement, achieving individual client accuracies of 0.77, 0.79, and 0.81, with the global model reaching an accuracy of 0.83. Moreover, the global model’s loss significantly reduces from 0.92 to 0.55 across five rounds, indicating successful convergence. These findings reinforce the framework’s efficacy in providing secure, decentralized anomaly detection within maritime networks. Future research will focus on incorporating neuromorphic architectures with FL to improve real-time performance in MASS communication.

Effective traffic management in smart cities demands robust link selection and reliable communication under dynamic conditions. To address these needs, we present a deep learning (DL)-based framework for intelligent link selection using LTE V2X communication data. Our approach utilizes LSTM, GRU, and 1D-CNN models to predict signal-to-noise ratio (SNR) patterns across multiple network operators. Among these, the GRU model achieved the highest performance with a 17.3% R2score, effectively modeling temporal variations in signal quality. The framework incorporates operator-specific optimization by accounting for signal strength, network congestion, and vehicular mobility patterns. Experimental evaluations demonstrate substantial improvements in both prediction accuracy and communication reliability compared to traditional approaches. Designed for edge deployment, the system supports real-time decision-making, making it highly suitable for adaptive traffic management in urban smart city environments.

The study develops an analysis framework built with deep learning techniques that extensively tests various architectures of modern Convolutional Neural Network (CNN) structures. The ResNet-18 model demonstrated the most successful implementation by reaching a training accuracy of 97.56% and validation accuracy of 86.92% at the same time. The corresponding training and validation losses amounted to 0.0754 and 0.6071, respectively. The LIME and Grad-CAM techniques of the XAI, and Occlusion were used to enhance the transparency of the model while providing components to better understand model decisions. The combination of accurate CNN models with interpretability tools produces successful explainable and robust classification in real-world application scenarios.

This paper proposes a new deep learning (DL) model for the detection of infant sleep disorders specific to Confusional Arousals (CA), Leg Restlessness (LR), and Sleep Apnea (SA) in 5G smart homes and e-Healthcare systems integration. The proposed framework, CHILD, consists of four critical layers: specific applications such as Infant Monitoring, Sensor and Data Acquisition, Artificial Intelligence systems and e-Healthcare and Smart Home systems. The smart home part increases the effectiveness of real-time environmental detection, the e-Healthcare system helps to provide convenient communication with doctors. The sleep disorder categorization problem is solved using an enhanced deep learning framework involving LSTM and GRU algorithms; the data are from sensors installed in the smart home., we obtained the 88% accuracy of LSTM model in consideration of the home automation and intelligent e-Healthcare system to enhance infant health and response actions.