Semantics in CNNs

This is the code for the paper "Interpreting the Information in Hidden Representations of Convolutional Neural Networks".

Adversarial Attacks

The adversarial sections of the paper consist of three files: adversarial.py, one_vs_two_adversarial.py, and adversarial_graphs.py.

The source images are expected to be a directory containing a directory corresponding to the class label, with source images of that class label inside that the attacks will be generated for. For example, here is a directory layout:

/AdversarialImages/
/AdversarialImages/accordion/
/AdversarialImages/accordion/ILSVRC2012_val_00004424.JPEG
/AdversarialImages/accordion/ILSVRC2012_val_00009350.JPEG
...

The map file is a pickle file containing the source classes along with the target classes to generate the adversarial attacks for. The target classes should include the correlation for the target class with the source class. The structure is as follows:

[
  [ 'accordion', ... ],
  [ 
    [ ('dog', 0.0370), ('cat', 0.0839), ... ],
    ...
  ]
]

The baseline concepts directory should follow the same format as the source images, but should contain class labels that have no overlap with either the source or target classes for the adversarial attack.

To generate the adversarial examples, run the following script:

> python adversarial.py \
    <location of source images> \
    <map file containing target classes> \
    --classes <one or more of the source classes to select (optional)> \
    --output-images <directory to write adversarial images (optional)>

The generate the 1 vs. 2 results, run one_vs_two_adversarial.py:

> python one_vs_two_adversarial.py \
    <location of the adversarial outputs from the last step> \
    <location of baseline concepts to compare against> \
    <map file containing target classes>

To generate the graphics for the results in the paper, run adversarial_graphs.py:

> python adversarial_graphs.py