Interpretable_CNNs_via_Feedforward_design

This is an implementation of the paper Interpretable Convolutional Neural Networks via Feedforward Design, maintained by Min Zhang and Jiali Duan.

Table of Content

• [Requirements]

  • Python3, keras, tensorflow, sklearn, cleverhans (Refer to https://github.com/tensorflow/cleverhans), pickle

• [Dataset] ( Hand-written digits classification)

  • [MNIST] ( train set: 60000, 28x28. Downloaded from Keras)
  • [CIFAR10] ( train set: 50000, 32*32. Downloaded from Keras)

• [Installation] (rensorflow, keras, pickle, sklearn and skimage)

  • [Sklearn Installation] Refer to http://scikit-learn.org/stable/install.html)
  • [Skimage Installation] (Refer to http://scikit-image.org/docs/dev/install.html)
  • [Optional: Jupyter Notebook] (Refer to http://jupyter.org/install.html)

• [Feedforward Steps]

  • Command python Getkernel.py, getting convolutional layers kernels
  • Command python Getfeature.py, getting feature after convolution
  • Command python Getweight.py, getting fully connected layers kernels and training accuracy
  • Command python mnist_test.py or python test.py, getting test accuracy

• [Adversarial attacks]

  • BP/ff models are provided for cifar10 and mnist dataset under folder dataset_structure_model
  • Models can be trained from scratch if no filename is specified
  • By changing adversarial attack methods, different algorithms can be tested
  • Refer to show_sample.ipynb to visualize generated adversarial samples

• [Example usage]

  • python cifar_keras.py -train_dir cifar_BP_model -filename cifar.ckpt -method FGSM
  • python cifar_keras.py -train_dir cifar_ff_model -filename FF_init_model.ckpt -method BIM

    cifar_keras.py:
    --batch_size: Size of training batches
    (default: '128')
    (an integer)
    --filename: Checkpoint filename.
    (default: 'FF_init_model.ckpt')
    --learning_rate: Learning rate for training
    (default: '0.001')
    (a number)
    --[no]load_model: Load saved model or train.
    (default: 'true')
    --method: Adversarial attack method
    (default: 'FGSM')
    --nb_epochs: Number of epochs to train model
    (default: '40')
    (an integer)
    --train_dir: Directory where to save model.
    (default: 'cifar_ff_model')

Contact me

Jiali Duan (Email: jialidua@usc.edu)
Min Zhang (Email: zhan980@usc.edu)