A REVIEW ON ADVERSARIAL ATTACKS AGAINST NEURAL NETWORKS AND THEIR DEFENSE METHODS

Abstract

Machine Learning (ML) is nowadays the core of technology and the main study area for different situations. It is indeed common to be the main subject of attackers at this kind of position. For instance, we can take as example the ML component of the last generation cars which have the autopilot functionality. Some would even think of damaging the ML component of the car system and try to perturbate the model to make the vehicle crash. In this case would be a real disaster, which of course a task at this level is being handled by car manufacturers. This situation is called usually as Adversarial Attack on Machine Learning Model. The necessity of protection against adversarial attacks is indeed crucial and thus we do have a variety of methods handling this. As attacks are so different to a ML model, defense methods are according to them. In this thesis study we will show main attack methods and their corresponding defenses. At the end we will show and compare the abovementioned methods effectiveness.