<![CDATA[Intrusion Detection Systems (IDS) based on machine learning face significant challenges in real-world deployment due to noise in network data and concept drift caused by evolving attack patterns. This study analyzes the robustness of three machine learning models Logistic Regression, Decision Tree, and Random Forest against noise and concept drift using the CICIDS2017 dataset. The experimental design includes baseline testing, noise robustness testing with three intensity levels (5%, 10%, 20%), and concept drift evaluation through temporal data splitting. Performance is measured using accuracy, precision, recall, and F1-score metrics, with robustness score calculated as a weighted average (40% baseline, 30% noise, 30% drift). Results show Logistic Regression achieves the highest robustness score (60.07%) due to excellent noise tolerance (1.09% F1-score degradation), followed by Random Forest (59.65%) and Decision Tree (57.60%). However, all models are categorized as NOT ROBUST against concept drift with degradation ranging from 29.64% to 30.34%, indicating sudden drift between Thursday and Friday data. The findings reveal that noise robustness and concept drift robustness are independent characteristics that do not correlate. Logistic Regression is recommended for practical deployment due to its optimal combination of robustness, interpretability, and computational efficiency. This research contributes to understanding model stability under non-ideal data conditions and emphasizes the necessity of implementing adaptive mechanisms such as periodic retraining and drift detection in operational IDS architecture.]]>
