Abstract:
Neural network usages are being popular in different sectors that make life easier and automated; so that security of neural network is a big concern. Against adversarial attack every Deep Neural Network are vulnerable. Adversarial attack is technique to specially design an image sample with adversarial noise. Several number of adversarial attack technique are found in recent research, which are fool neural network with high misclassify accuracy. There is also various defense mechanism are proposed and build with Deep Neural Network to defend and increase robustness of the main classifier neural network model. However, there are very few model can work with high resolution image data and work with pre-trained neural network classifier. The main objective of this research was to proposed and developed a model which can integrate with any existing trained Neural network and more generic defense against adversarial attack tool called ADDA-Adv. Based on proposed model detect highly distorted image and reject those sample to avoid misclassification. Additionally, this work is intended to work with high resolution adversarial image sample. ADDA-Adv. tool restore the adversarial sample and provide 89.23 percent accuracy.
