A Domain-Adaptive Gated Deep Learning Framework with Dynamic Dropout Optimization for Network Intrusion Detection System

Due to the rapid growth of the modern network infrastructures and the rise in the sophistication of the attacks by criminals based on networks, intrusion detection
system (IDS) has become a crucial component in offering network security. The common machine learning and the existing deep learning-based techniques might be unable
to deal with the extremely dynamic traffic dynamics, and lead to such issues as poor generalization, excessive false alarms, and poor detection of new and evasive attacks.
The solution to these challenges is solely to possess smart, adaptive and tough detection frameworks that can learn sophisticated representations across different
environments. A new intrusion detecting framework is presented in this paper that is based on a combination of a Domain-adaptive Gated Deep Belief Network (DomG�DeNet) and an enhanced optimization method known as Builder-on-Zebra Recurrent Dropout Optimization (BoZ-RDO). The suggested DomG-DeNet is an improvement on
feature abstraction and facilitates the adaptation of the domain, thus letting the model extrapolate across several benchmark datasets. At the same time, BoZ-RDO
dynamically controls the dropout rates during training, which is based on biological processes, to control the complexity of the model and prevent overfitting. The framework
implies the massive preprocessing and extraction of the features on the familiar datasets, which are ISCX-IDS2012, CICIDS2017, and CSE-CIC-IDS2018, and the following
multi-class attack classification and the training of deep models in the most optimal way. It is also demonstrated in experiments that the proposed solution is significantly
more effective than the more recent models such as CNN-BiLSTM, Transformer, and GRU-Attention. It achieves an accuracy of 99.1%, precision of 99.0%, recall of 99.2%,
F1-score of 99.1%, and an AUC-ROC of 99.5%. These findings validate the usefulness, scalability, and strength of the suggested framework in the real-time and large-scale
intrusion detection situations.
Keywords: deep learning; domain adaptation; dynamic dropout optimization; intrusion detection system; network security