Physical Sciences and Mathematics Commons^™

Semantic Adversarial Attack On Support Vector Machine, Yessica Rodriguez

Theses and Dissertations

Despite the breakthroughs in machine learning, most classifiers are not robust against adversarial attacks. They can be easily fooled by adversarial examples. These examples can be created in a variety of ways. In this thesis, the ideas of detecting edges or critical pixels in an image are investigated that could be used for fooling classifiers. Identifying those critical pixels in an image can lead the way to fix the vulnerabilities and thus making it robust against cyber-attacks. For testing, a Support Vector Machine (SVM) is used to see the success of the adversarial examples generated.