This repository contains the code for Deviations in Representations Induced by Adversarial Attacks.
Reported running times are approximate, intended to give a general idea of how long each step will take. Estimates are based on times encountered while developing on Ubuntu 22.04 with hardware that includes an AMD Ryzen 9 3950X CPU, 64GB of memory, and an NVIDIA TITAN RTX GPU with 24GB of memory. The intermediate results utilize about 29 gigabytes of storage.
The code was developed using Python 3.10 on Ubuntu 22.04. Other systems and Python versions may work, but have not been tested.
Python library dependencies are specified in requirements.txt. Versions are pinned for reproducibility.
- Optionally create and activate a virtual environment.
python3 -m venv env
source env/bin/activate- Install Python dependencies, specified in
requirements.txt.- 4 minutes
pip3 install -r requirements.txtBy default, output is saved to the ./workspace directory, which is created automatically.
- Train a ResNet classification model.
- 1 hour
python3 src/train_net.py- Evaluate the model, extracting representations from the corresponding data.
- 1 minute
python3 src/eval_net.py- At each layer, calculate pairwise distances between representations, for normalization.
- 2 hours
python3 src/calc_pairwise_distances.py- Adversarially perturb test images, evaluating and extracting representations from the
corresponding data.
- 7 hours
python3 src/attack.py- Calculate the distances between representations for original images and their adversarially
perturbed counterparts.
- 3 minutes
python3 src/calc_distances.py- Analyze data by first normalizing and then plotting and tabulating.
- 3 seconds
python3 src/analyze.py@misc{steinberg2022deviations,
doi = {10.48550/ARXIV.2211.03714},
url = {https://arxiv.org/abs/2211.03714},
author = {Steinberg, Daniel and Munro, Paul},
title = {Deviations in Representations Induced by Adversarial Attacks},
publisher = {arXiv},
year = {2022},
eprint = {2211.03714},
archivePrefix = {arXiv},
primaryClass = {cs.LG}
}