Research on Target Detection and Anti-Attack Method Based on Neural Network

Abstract

The well trained neural networks have performed much better than human beings in many daily task, like object detection and image classification. The security has also been an important issue while neural network makes our daily life much more convinient. Adversarial attack can attack on the input with small perturbation to let network make wrong prediction. Based on the information that attacker can get during the attack, we can divide the attack methods into white box attack and black box attack. Currently, all the black box attack methods are designed for image classification, and there is no black box attack for object detection. In this paper, we successfully apply the Boundary Attack which was originally designed for image classification on object detection task.