DHPAN: A Real-Time Adaptive Deep Learning Framework for Streaming Intrusion Detection

The growing number and complexity of network traffic in contemporary cloud and IoT systems require dynamic and real-time systems of intrusion detection. In this paper, a multimodal deep learning framework known as Dynamic Hyper-Parallel Attention Network (DHPAN) is presented, which is used to detect intrusion by streaming. The DHPAN combines a Lightweight Feature Embedding Layer (LFEL) to support compact representations, a Convolutional Recurrent Neural Network (CRNN) with Anomaly-Aware Attention (AAA) to both learn spatial-temporal and contextual patterns and a Hyper-Parallel Dual Optimization (HPDO) strategy that incorporates both Adaptive Adam and Evolutionary RMSprop. The framework applies seven stages of processing network traffic, such as preprocessing and adaptive feature selection, interpretability, and evaluation. The experimental findings on NSL-KDD and CIC-IDS-2017 data sets show a higher performance, with accuracy of 99.59% and 99.68%, respectively. Measures made at network level show low detection latency, high throughput and low packet drop, which confirm real-time applicability. Ablation and threshold sensitivity tests validate the role of every module and the most suitable selection of anomaly score. DHPAN provides proactive zero-day and multi-vector attack response when preserving network QoS, which provides a high-quality, scalable and understandable intrusion detection platform on next-generation networked systems.