Coursework Project for Computer Science: Advanced Topics
With the wide use of computer vision technologies, the security problem of such technologies are increasingly concerned by scientists. Recently, several image fooling methods have been developed, which confuse the neural network with minor perturbation. Meanwhile, ODENet, a new kind of neural network, has been proposed in NeurIPS 2018 and has achieved the best paper of the conference. This new network shows great potential against image fooling, which arouses our study interest and inspires us to propose a new model named Random ODENet against image fooling. The basic idea is to introduce randomness in the structure of ODENet. For more details about this work, please see our course work report and introduction ppt in this repository
- Follow the installation guide from this link to install ODENet.
- Clone this repository.
- Download the MNIST dataset in "fooling/mnist", and download 2 pretrained models from this link, and add it to "fooling" folder.
- Download foolbox
- You can then reproduce our experiment result by running "fooling/mnist_fooling.py", "fooling/ode_fooling.py", "fooling/ode_fooling_random.py".