Analysis of evasion attack defense methods in text classified training dataset

Abstract

Classification algorithms built different kind of feature representations based on training datasets. The major threat on training datasets are, they affected by various attacks. The unstructured training datasets are faced the challenges when they convert into structured datasets. The tiny text perturbation in the original training dataset will cause misclassification and incorrect predictions in machine learning. The different classification algorithms measurements help to detect the evasion attack on training dataset. To compare different defense methods helps the way of mitigating training dataset attacks. The experimental results prove that the text classifier training dataset secured from the evasion attack.