A New Poisoning Attacks on Deep Neural Networks

Abstract

In healthcare field, many machine learning schemes have been applied in analyzing image content dataset. Among them, deep neural networks (DNNs), also known as deep learning, catches much attention. However, if deep neural networks are compromised by the attacker, medical diagnosis may be wrong, which leads to vital result. Recently, we find a new poisoning attack on DNNs may possibly happens due to poisoning dataset. This new poisoning attack, Category Diverse attack, has better ability to paralyze DNNs. Our performance experiments show our Category diverse attack actually leads to large accuracy drop of DNNs. We hope this discovery can help the information experts can improve the medical dataset quality in the future.