Enhancing Adversarial Attack Detection: Insights from Ensemble Learning, Deep Dyna-Q, and VARMAx Models

Abstract

Machine learning systems are increasingly crucial for various applications but are also vulnerable to hostile attacks. Current methods are often inflexible and imprecise. This paper presents a sophisticated ensemble model that combines Deep Dyna Q Learning and VARMAx procedures with Active Machine Learning Adversarial Attack Detection frameworks. The model uses a dynamic and responsive methodology, combining data pretreatment methods like tokenization and numerical conversions. Techniques like CNN, SVM, Random Forest, XGBoost, and Logistic Regression are used to improve detection capabilities. The model incorporates uncertainty sampling and query-by-committee for adaptability to new adversarial tactics. VARMAx operations improve prediction accuracy, while Deep Dyna Q Learning anticipates attack vectors. The system's defensive mechanisms are strengthened with GridCAM++ for explainability and GAN-based sample generation. The model outperforms current approaches in precision, accuracy, recall, and AUC, and reduces detection delays