ART for few shot learning

[anukriti15]

Can we use the evasion attacks for few-shot learning models? Forward function is the problem as support and query set is involved

[beat-buesser]

Hi @anukriti15 My first thought would be that it should be possible. Are you looking for white- or black-box attacks?

[anukriti15]

My preference will be white box attacks

[beat-buesser]

For white-box attack you will have to create a new estimator specific for your model that internally loads the support set and applies it appropriately in estimator methods like loss_gradient, predict, etc. For black-box attack you might be able to use the existing estimator BlackBoxClassifier and include the support set in the custom function predict_fn.

[anukriti15]

Thanks a lot. I will attempt this and get back to you for any further queries.

[anukriti15]

Also, does ART support adversarial attacks for text or tabular data?

[beat-buesser]

There are capabilities for tabular data with continuous features.