Understanding Adversarial Training: Improve Image Recognition Accuracy of Convolution Neural Network

Traditional methods of computer vision and machine learning cannot match human performance on tasks such as the recognition of handwritten digits. Recently many researchers work on Convolution Neural Network for image recognition, and get results as good as human being. Additionally, Image recognition task is getting more popular and high demand to apply to other fields, but also there are still many problems to utilize in everyday life. One of these problems is that several machine learning models, including neural networks, consistently misclassify adversarial examples—inputs formed by applying small but intentionally worst-case perturbations to examples from the dataset, such that the perturbed input results in the model outputting an incorrect answer with high confidence. The main purpose of this thesis is to use Convolution Neural Network (CNN) as a tool to recognize and classify images in 4 types of data sets; MNIST (hand-writing digits), CIFAR10 (animal, food, vehicle pictures), MNIST and CIFAR10 adversarial example. The optimal performance on MNIST and CIFAR10 was achieved by using two essential steps. First, we created a basic convolutional neural network; single layer, defined hyper-parameters in Keras, then train and test by the datasets and computed accuracy and loss of recognition. Second, I modify the network to adjust a network structure and hyper–parameters one by one, then compare to the basic network. Next, I found out the optimal performance network on MNIST and CIFAR10 adversarial examples by these two steps. First, I attached the MNIST and CIFAR10 by making them slightly different, and put these datasets to the networks which were adjusted already. I got result and compare to best accuracy. Second, I change a training and predicting process to adapt the adversarial example. The idea is to train the same CNN for several times (odd numbers of time), obtaining different weights in different times. And putting test data into these CNNs, then the final result is determined by voting. Using this approach to provide examples for adversarial training, I reduced the test set error of the network on the MNIST and CIFAR10 dataset