Document worth reading: “Adversarial Examples in Modern Machine Learning: A Review”

Recent evaluation has found that many households of machine learning fashions are weak to adversarial examples: inputs which could be notably designed to set off the aim model to produce defective outputs. In this survey, we take care of machine learning fashions in the seen space, the place methods for producing and detecting such examples have been most extensively studied. We uncover numerous adversarial assault methods that apply to image-space content material materials, precise world adversarial assaults, adversarial defenses, and the transferability property of adversarial examples. We moreover deal with strengths and weaknesses of various methods of adversarial assault and safety. Our aim is to supply an in depth safety of the sector, furnishing the reader with an intuitive understanding of the mechanics of adversarial assault and safety mechanisms and enlarging the group of researchers discovering out this elementary set of points. Adversarial Examples in Modern Machine Learning: A Review