Effects of Selection Bias on Online Adversarial Aware SVM When Facing an Evasion Attack

In evasion attacks, the success of machine learning (ML) depends on its ability to detect an attack in an adversarial setting. It is important that these ML models are trained regularly to keep abreast with evolving evasion attacks. We present an investigation of the effects of selection bias and class imbalance in training an Adversary-Aware Online SVM (AAOSVM). We show that incorrect samples can compromise the model's ability to detect attacks. Comparison results showed better True Positive Rates (TPR) performance than the Online Support Vector Machine (OSVM), and that AAOSVM was sensitive to selection bias but not to class imbalance.