DLproject

A project about data poisoning on satellite images.

Get the data

Data is not directly available here as it's too big for github. However you can download it here:

wget https://zenodo.org/records/7711810/files/EuroSAT_RGB.zip?download=1

How to run

In order to train the models, just use the train.py script:

python3 train.py --models cnn resnet googlenet

ViT (vision transformer) is listed as an option but not currently working yet.

The training parameters are all contained in config.py and should only be modified there.

Data poisoning part

Check here for now: i will integrate the repos later. Run the data poisoning experiment with:

python3 brew_poison.py --dataset EUROSAT 

It could happen that the program complains about Local Cache (expecially if running on google colab). In that case, reinstall the dataset library: pip install -U datasets. Running this complete experiment takes approximately 2 hours.

References

Dataset: EUROSAT Types of attacks

• Gradient Matching
• Bullseye Polytope Architectures:
• ViT