This GitHub repository contains the official code for the paper,

Transferability of features for neural networks links to adversarial attacks and defences
Shashank Kotyan, Moe Matsuki and Danilo Vasconcellos Vargas
PLOS One (2022).

If this work helps your research and/or project in anyway, please cite:

@article{kotyan2022transferability,
  title={Transferability of features for neural networks links to adversarial attacks and defences},
  author={Kotyan, Shashank and Matsuki, Moe and Vargas, Danilo Vasconcellos},
  journal={PloS one},
  volume={17},
  number={4},
  pages={e0266060},
  year={2022},
  publisher={Public Library of Science San Francisco, CA USA}
}

The code is tested on Ubuntu 18.04.3 with Python 3.7.4.

To run the code in the tutorial locally, it is recommended,

• a dedicated GPU suitable for running, and
• install Anaconda.

The following python packages are required to run the code.

• matplotlib==3.1.1
• numpy==1.17.2
• seaborn==0.9.0
• tensorflow==2.1.0

1. Clone the repository.

git clone https://github.com/shashankkotyan/RepresentationMetrics.git
cd ./RepresentationMetrics

2. Create a virtual environment

conda create --name rm python=3.7.4
conda activate rm

3. Install the python packages in requirements.txt if you don't have them already.

pip install -r ./requirements.txt

4. Train and evaluate a normal or a Raw Zero-Shot classifier.

python -u code/run.py [ARGS] > run.txt

TBD

• To evaluate the DBI and AM metrics with the adversarial examples. Please generate the adversarial examples using the repository Dual Quality Assessment

• Tutorials
• Addition of Comments in the Code
• Cross Platform Compatibility
• Description of Method in Readme File

Representation Metrics is licensed under the MIT license. Contributors agree to license their contributions under the MIT license.

TBD

You can reach me at shashankkotyan@gmail.com or @shashankkotyan. If you tweet about Representation Metrics, please use the following tag #raw_zero_shot, and/or mention me (@shashankkotyan) in the tweet. For bug reports, questions, and suggestions, use Github issues.