Dynamic Feature Extraction in Streaming Data for Real-Time Analytics

Abstract

The explosive increase in traffic of networkcreatedby means ofInternet of Things (IoT) devices, industrial processes, and web applications has complicated cybersecurity, rendering traditional approaches obsolete due to their high computational costs and delays, and their lack of adaptability to emerging traffic patterns. Conventional techniques suffer from poor discrimination, and they fail to achieve high detection accuracies. To address these limitations, the proposed research introduces a Seven-Spot Ladybird–mutated Deep Recurrent Neural Network (SSL-DRNN), in which a DRNN is employed to capture sequential dependencies, while the SSLO algorithm is utilized for optimal feature selection. The model is evaluated using the Network Intrusion dataset from Kaggle, which containsmalicious andnormal network traffic samples 28,30,743. All input traffic data are normalized through Z-score method to achieve unit varianceandzero mean, and relevant features are further extracted using Incremental Principal Component Analysis (IPCA).The experiments prove that the approach performs well in separating normal traffic from malicious traffic, which achieves very high values of F1 score of 98.5%, accuracy of 98.4%, and False Positives Rate (FPR) of 0.012. The implementation of the model is done through the use of Python code written using TensorFlow, Keras, Scikit-Learn, NumPy, and Pandas. The suggested SSL-DRNN method greatly improves the anomaly detection in network traffic, which helps in classification accuracy, decreasing false positives, and increasing cybersecurity performance and reliability.