Hybrid Model for Imbalance Correction in Intrusion Detection Systems Using Advanced Optimization Techniques and Graph Neural Networks

Intrusion detection systems regularly battle with class disproportions in datasets steering to weakened accuracy detection. This paper uses SMOTE to define the challenge and also incorporates modern machine learning models such as LSTM and GNN for improved identification of anomaly. The UNSW-NB15 dataset is used to pre-process steps like PCA, normalization, and imbalance correction techniques. These techniques are used to increase representation of data. The hybrid technique demonstrates an accuracy of 80%, precision above 75%, and AUC transcending 80% surpassing conventional methods in detecting both normal and abnormal activities. LSTM transcended in seizing temporal patterns, and GNN handles relational data improving robustness for detection. The average training time of 95 seconds defines the computational efficiency of system. These results show the flexible and real-world relevance of the framework in defining cybersecurity threats. The future work will focus on ensemble learning methods to farther enlarge identification capacities paving the way for more strong intrusion detection systems.