shu-hai / sa_dnn

Sensitivity Analysis of Deep Neural Networks (AAAI-19 paper)

SA_DNN

This is the repository for the following AAAI-19 paper:

Shu, H., and Zhu, H. (2019) Sensitivity Analysis of Deep Neural Networks. The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI-19), pp. 4943-4950 [DOI]

Also see the talk slides.

We also proposed a method, called mFI-PSO, for Adversarial Image Generation based on the Manifold-based First-order Influence (mFI) measure introduced in the paper. See https://github.com/shu-hai/mFI-PSO for details.

Instructions:

Use CIFAR10_sample.py and MNIST_sample.py to obtain the CIFAR10 and MNIST datasets.

ResNet50.py and DenseNet121.py are the two networks, which are called to be trained by CIFAR10_ResNet50.py, CIFAR10_DenseNet121.py, MNIST_ResNet50.py or MNIST_DenseNet121.py.

Then for the two benchmark datasets, take CIFAR10 and DenseNet121 for example. Run CIFAR10_DenseNet121_IF_setupX.py for Setup X in the paper, where X=1,2,3,4. To summarize the results, first use the R code CIFAR10_DenseNet121_IF_setupX_result.R, and then use the python code CIFAR10_DenseNet121_IF_setupX_result.py for the plots.

sa_dnn's Issues

some questions

Could you tell me what the "set_part" means? Thanks~

How to change the objective function?

Dear Dr. Hai Shu,

I would like to use your metric in my future projects.
But I cannot figure out where the objective function (e.g. Cross Entropy) is implemented in the codes. It seems to me that the objective function is already implemented in grad_f objects [as nabla_f(omega0) regarding your original article]. Can you help me clear up my confusion?

Thank you very much

Best regards