Anomaly Detectionin Network Traffic Scenarios by Resampling and Majority Voting with Concept Drift: A Hybrid Approach

Addressing data imbalance is a critical difficulty in the setting of a heavily skewed dataset because the dominant class disproportionately influences classifier accuracy, especially when inadequate data impede the learning of minority class characteristics. For multi-class cases, traditional binary classification approaches are inadequate. The suggested architecture uses a Weighted Majority Voting Classifier (WMVC), noise cleaning, limited under-sampling, and oversampling to overcome this. One important step is to set an average class size limit by dividing the entire sample number by the number of classes. One or more majority classes receive Random Under sampling, all minority classes receive an up sample using the Adaptive Synthetic Sampling Approach, and noise reduction is achieved via Tomek Link. The resulting WMVC is contrasted with Convolutional Neural Network (CNN) classifiers, XGBoost, and six additional ensemble techniques by utilizing tree-based ensemble algorithms. Comparative study shows that balanced data performs better than unbalanced data. WMVC significantly beats CNN, XGBoost, and the other ensemble methods, and it improves performance for difficult minority classes while successfully decreasing classifier bias towards the majority class.