Mitigations for the attacks

[dsvrsec]

Are there any mitigations available for NLP models to defend against these attacks or make the model robust.I am thinking Adversarial training can be one.

[jxmorris12]

Hi @kumarvrsec – right now we support adversarial training via concatenating and replacing the dataset with perturbed examples. This is via textattack train with the --attack argument. We're planning on adding more defense techniques in the next few months.

[dsvrsec]

Hi,Are there any defense(mitigations) techniques that are added to textattack or available in the literature

[jxmorris12]

Hi @dsvrsec -- we enable adversarial training via textattack train --attack.