Examples

There are multiple examples available for Dioptra that demonstrate its capabilities across different models, datasets, adversarial attacks, and adversarial defenses. Here you will find instructions on how to prepare your environment for running an example and a table that lists what is available.

Setup

To prepare your environment for running the examples, follow the linked instructions below:

1. Clone the repository and build the included containers

2. Apply the provided cookiecutter template and run Dioptra

3. Create and activate the Python virtual environment for the Dioptra examples and download the datasets using the download_data.py script

4. Edit the docker-compose.yml file to mount the data folder in the worker containers.

5. Initialize and start Dioptra

6. Register the custom task plugins for Dioptra’s examples and demos

7. Register the queues for Dioptra’s examples and demos

8. Start JupyterLab and open the demo Jupyter notebook (ipynb file extension)

Steps 1–3 and 6–7 only need to be run once. Returning users only need to repeat Steps 4 (if you stopped Dioptra using docker compose down) and 8 (if you stopped the jupyter lab process).

List of Examples

The current list of examples for Dioptra is provided in the table below. It is recommended that newcomers start with the Tensorflow MNIST Classifier example.

Name	Library	Models	Dataset	Attacks	Defenses	GPU
Tensorflow MNIST Classifier	Tensorflow	ShallowNet, LeNet	MNIST	FGM		No
Tensorflow MNIST Pixel Threshold	Tensorflow	ShallowNet, LeNet	MNIST	Pixel Threshold		No
Tensorflow MNIST Model Inversion	Tensorflow	ShallowNet, LeNet	MNIST	Model Inversion		No
Tensorflow MNIST Feature Squeezing	Tensorflow	ShallowNet, LeNet	MNIST	FGM, Feature Squeezing, CW, Deepfool, JSMA		No
Tensorflow Backdoor Poisoning	Tensorflow	LeNet	MNIST	Backdoor Poisoning	Spatial Smoothing, JPEG Compression, Gaussian Augmentation	No
PyTorch MNIST Membership Inference	PyTorch	LeNet	MNIST	Membership Inference		No
Tensorflow ImageNet ResNet50 Demo	Tensorflow	ResNet50	ImageNet	FGM, Pixel Threshold	(FGM only) Spatial Smoothing, JPEG Compression, Gaussian Augmentation	Yes
Tensorflow Adversarial Patch Demo	Tensorflow	LeNet (MNIST), VGG16 (Fruits 360), ResNet50 (ImageNet)	MNIST, Fruits 360, ImageNet	Adversarial Patch	Spatial Smoothing, JPEG Compression, Gaussian Augmentation, Adversarial Training	Yes (Fruits 360, ImageNet only)
PyTorch Detectron2 Demo	PyTorch	RetinaNet	Road Signs	Backdoor Poisoning		Yes