Detection of data-driven discovery attacks on machine learning classifier algorithms

Recently, websites have become very smart in order to meet user needs and make their lives easier. Such that systems that can perform artificial intelligence and machine learning processes have started to be on websites. Currently, web servers that can classify with machine learning can serve users. When these systems were designed, there was no security concern and they were developed without considering the security issue. Therefore, these systems are vulnerable to attack. But it is impossible to ignore cyberattacks today. In this study, in order to reveal the weaknesses of the classification system, we first obtained a limited amount of exploration data about information the system by attacking the system like an adversary. Then, we prepared the obtained data to attack the system by transforming it into a larger data set in our system, which we trained with gradient descent method. The results were obtained by making attacks on the classification system with the attack dataset. When the results obtained were compared with the actual results, it was ensured that the system was directed to wrong side and the classification performance decreased as a result of the attacks.