Exploring the Effect of Dimensionality Reduction Techniques on Filtration Attacks

This paper investigates the effectiveness of dimensionality reduction techniques combined with stratified sampling on the performance of detection models for (in/ex) filtration attacks. Utilizing a Random Forest model, we explore the impact of reduced feature sets derived by dimensionality reduction methods, namely PCA, ICA, AE, and GP. The evaluations are performed on three cybersecurity datasets across various granularity levels. Results indicate that while models trained on reduced feature sets perform comparably to those trained on full feature sets, they offer significant advantages such as reduced computational demands and enhanced explainability, suggesting potential for operational efficiencies in cybersecurity.