Robust Architecture Search

This github repository contains the official code for the papers,

Towards Evolving Robust Neural Architectures to Defend from Adversarial Attacks
Shashank Kotyan and Danilo Vasconcellos Vargas
GECCO (2020).

Evolving Robust Neural Architectures to Defend from Adversarial Attacks
Shashank Kotyan and Danilo Vasconcellos Vargas,
AISafety Workshop (2020).

Is Neural Architecture Search A Way Forward to Develop Robust Neural Networks?
Shashank Kotyan and Danilo Vasconcellos Vargas
JSAI (2020).

IMPORTANT

In the current version of the code, the robustness evalution is not implemented in the code. We are currently testing for the compatibility and reproducibility of the adversarial examples and results before adding it in the repository to ensure the quality of the code remains intact.

In the meantime, one can generate their own adversarial samples using the repository of Dual Quality Assessment and make necessary changes in fitness evaluation of the evolved models in run_model function in worker.py file. Fitness of the evolved models is calculated in the line 113

Citation

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

@inproceedings{kotyan2020towards,
  title={Towards evolving robust neural architectures to defend from adversarial attacks},
  author={Kotyan, Shashank and Vargas, Danilo Vasconcellos},
  booktitle={Proceedings of the 2020 Genetic and Evolutionary Computation Conference Companion},
  pages={135--136},
  year={2020}
}

@inproceedings{kotyan2020evolving,
  title={Towards evolving robust neural architectures to defend from adversarial attacks},
  author={Kotyan, Shashank and Vargas, Danilo Vasconcellos},
  maintitle = {International Joint Conference on Artificial Intelligence - Pacific Rim International Conference on Artificial Intelligence (IJCAI-PRICAI)},
  booktitle = {Workshop on the AISafety},
  year={2020}
}

@inproceedings{kotyan2020neural,
  title={Is Neural Architecture Search A Way Forward to Develop Robust Neural Networks?},
  author={Kotyan, Shashank and Vargas, Danilo Vasconcellos},
  booktitle={人工知能学会全国大会論文集 第 34 回 (2020)},
  pages={2K1ES203--2K1ES203},
  year={2020},
  organization={一般社団法人 人工知能学会}
}

Testing Environment

The code is tested on Ubuntu 18.04.3 with Python 3.7.4.

Getting Started

Requirements

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.

• GPUtil==1.4.0
• matplotlib==3.1.1
• networkx==2.3
• numpy==1.17.2
• scipy==1.4.1
• tensorflow==2.1.0

---

Steps

1. Clone the repository.

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

2. Create a virtual environment

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

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

pip install -r ./requirements.txt

4. Run the Robust Architecture Search Code with the following command.

python -u run_evolution.py [ARGS] > run.txt

5. Calculate the statstics for the evolution.

python -u run_stats.py > run_stats.txt     

Arguments to run run_evolution.py

TBD

Notes

• It is recommended to run the code on a multi-gpu system to ensure faster evolution. However, changing the number of workers in num_workers.txt to 1 will ensure the evolution on a single GPU system.
• Setting num_workers.txt to number of GPUs your system has will run the code optimally utilising maximum performance by the GPUs.

Milestones

• Include Robustness Evaluation
• Toy Example for Evolutionary Strategy
• Addition of Comments in the Code
• Cross Platform Compatibility
• Description of Method in Readme File

License

Robust Architecture Search is licensed under the MIT license. Contributors agree to license their contributions under the MIT license.

Contributors and Acknowledgements

TBD

Reaching out

You can reach me at [email protected] or @shashankkotyan. If you tweet about Robust Architecture Search, please use the tag #RAS and/or mention me (@shashankkotyan) in the tweet. For bug reports, questions, and suggestions, use Github issues.