🚧 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

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

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)