🚧 project is in an undefined stage of development, come back in July (of your favorite future year) for hopefully functional version 🚧

Evolutionary GENerated Adversarial examples

Štěpán Procházka (author, Charles University in Prague)
Roman Neruda (supervisor, Academy of Sciences of the Czech Republic)

Table of Contents

0. Introduction
1. 10 minutes to Evgena
2. 100 minutes to Evgena
3. Features

10 minutes to Evgena

clone the repository

$ git clone https://github.com/proste/evgena.git
$ cd evgena

create environment, activate and install requirements

$ python3 -m venv .env
$ . .env
$ pip install -r requirements.txt

run jupyter notebook

$ jupyter notebook

100 minutes to Evgena

Features

• datasets

  • fashion MNIST
  • MNIST
  • CIFAR-10
  • CIFAR-100
  • (imagenet)

• approaches

  • gradient based methods on target model (kind of baseline)
  • gradient based methods on surrogate model
  • GA methods
  • joint GA and gradient based methods on surrogate model

• tasks (each task with and without constraint on visual similarity)

  • given input example, modify to get desired class
  • given input examples, find universal modification to get desired class
  • given target class, generate corresponding input (inverse mapping)

• performance indicators

  • inference count on target model
  • (time/space) complexity
  • "aesthetics" of results
  • pros/cons in terms of constraints forced on environment (model, data)