A Stacking Ensemble Framework for Robust Detection of Data Poisoning Attacks

Machine learning (ML) technologies have become foundational in critical domains such as cybersecurity, healthcare, and autonomous systems. However, their increasing reliance on large-scale training data exposes them to data poisoning attacks—where adversaries manipulate training inputs to degrade model performance. Such attacks can cause misclassifications with severe consequences, particularly in high-stakes environments. This study introduces a robust stacking ensemble framework for the detection and mitigation of data poisoning attacks. The proposed model integrates four supervised classifiers—K-Nearest Neighbors (KNN), Random Forest (RF), Decision Tree (DT), Gradient Boosting (GB), and Logistic Regression (LR)—within a meta-learning architecture to enhance detection accuracy and resilience. We evaluate the model on four benchmark cybersecurity datasets: UNSW-NB15, BotDroid, CTU-13, and CICIDS-2017, representing diverse attack scenarios and feature distributions. Poisoning is simulated at six intensities (0% to 25%) by injecting mislabeled or adversarial perturbed data. The ensemble model consistently outperforms individual classifiers across all datasets and poisoning levels. Notably, it achieves 99.56% accuracy and a 99.69% F1-score on CICIDS-2017 with 25% poisoning, and maintains 95.58% accuracy on the BotDroid dataset, where baseline models degrade significantly.