Improving Neural Networks Robustness for Computer Vision
Year: 2018
Type: dissertation
Abstract: Image classification had been a challenging task for computers until neural networks reinvented and achieved the state-of-the-art in a competition in 2012. This has changed the research direction of the computer vision tasks and many other fields. However, with all the success of neural networks in image classification, state-of-the-art networks are still vulnerable to major challenges. The first challenge is classifying blurred images, which can be caused when a camera is unfocused, during motion or using image processing filters. The second challenge is classifying affinely transformed images such as rotation, shitting, or scaling. This problem is neglected because the effects of transformation are not obvious as much as other problems. The third challenge is known as adversarial examples which are images that include structured gradient, could be shown as noise for humans or not shown at all, generated from neural networks. The challenge with adversarial images is that they are imperceptible and misclassified with higher confidence rate. This dissertation introduces methods to make neural networks more robust to these challenges and using a new architecture. One of the architecture is named a Simultaneous Convolutional Neural Network (SCNN) that aims to make a model more robust by train a pair of sub-networks. The sub-networks produce learning for original and perturbed (blurred or affine transformed) images at the same time. During training, the model learns the difference between the two input images to make a model robust against blurring and affine transformations. The first and second contributions are based on SCNN architecture to overcome irregularity when applying data augmentation during training. The first contribution overcomes the reduction of performance of classifying high-quality images when the training dataset includes low-quality images. The second contribution enhances the classifying performance using affine transformations such as in data augmentation. The third contribution presents an enhancement of recovering the misclassified infected images of some types of adversarial examples. The method Implicitly applies to blur on the adversarial images, which results in a high reduction in perturbations of adversarial images. The role key in this contribution is explicit of a model that robust to blurred images, as in the first contribution, to classify the resulted images after blurring adversarial examples. The last contribution analyzes the features of the adversarial images and how they affect original images. We hypothesis that these images have different features that affect the output feature maps of a neural network. The proposed solution isolate adversarial examples from clean images instead of forcing a model to train and classify them within original classes. Our architecture successfully classify the new adversarial images even when the attacker knows the architecture and weights of the model.
Author: Faris Kateb
Institution: University of Colorado Colorado Springs
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: King Abdulaziz University
OpenAlex ID: W2901588101
