An Impact of Evasion Attacks on Machine Learning Algorithms
Lourdu Mahimai Doss P, Muthumanickam Gunasekaran · 2024
The research aims to demonstrate the susceptibility of Machine Learning Algorithms to adversarial attacks. Machine Learning is pivotal in diverse applications across homogeneous and heterogeneous environments, including Computer Vision, Health Care, Stock Markets, Banking, and Intrusion Detection Systems. The study involves training machine learning classifier models using randomly generated datasets and assessing the accuracy of each classifier. Subsequently, the research introduces perturbations during the test phase using the projected gradient descent algorithm and the projected gradient descent line search algorithm. The injection of these perturbations results in a decrease in accuracy for various machine learning classifiers. The impact of Evasion attacks on classifier models is quantified, revealing substantial decreases in accuracy for different classifiers. Specifically, the accuracy of SVM Linear decreases from 46.00% to 10.70%, SVM RBF from 91.20% to 08.00%, Logistic from 54.00% to 16.00%, kNN from 88.00% to 11.20%, Decision Tree from 86.80% to 12.00%, and Random Forest from 88.40% to 12.34%. This comparative analysis underscores the success of Evasion attacks in compromising the accuracy and integrity of machine learning classifiers.