Using Backdoor Attacks to Find Areas of Interest in Images

The power of neural networks increases in popularity more and more everyday. However, as for any powerful and popular tool, the security is hard to gain, here it is also the case. In recent years, backdoor attacks have become more and more popular through malicious parties, being risky for both organisations and people. In this paper we discuss the main methods of prevention against this type of attack. We also provide an experiment, with the help of which we tried to see what part of an image influences its classification. Therefore, using an algorithm for identifying backdoor attacks we made an experiment intended to find the interest zone in X-ray images.