search

QData / TextAttack

TextAttack 🐙 is a Python framework for adversarial attacks, data augmentation, and model training in NLP https://textattack.readthedocs.io/en/master/
https://textattack.readthedocs.io/en/master/
MIT License
2.87k stars 383 forks source link

Add new attack recipe #385

Closed RishabhMaheshwary closed 2 years ago

RishabhMaheshwary commented 3 years ago

Request to add a new attack recipe: Generating Natural Language Attacks in a Hard Label Black Box Setting, accepted in AAAI 2021 main conference. Paper: Link to paper. Code: Link to code.

Brief Outline: The paper proposes a novel decision based attack strategy that utilizes population-based optimization algorithm to generate adversarial examples by observing only the topmost predicted label.

jinyongyoo commented 3 years ago

Hi @RishabhMaheshwary, thank you sharing the paper. If you don't mind, could you make a PR for the new attack recipe? Please feel free to create any new components (e.g. search method, transformation, constraint) that is needed for your attack.

While most of the recipes have been added by TextAttack developers, we also would like to have some community input and feedback. If you have any questions or issues please let me know! You can either create Github issues or directly ask questions in our slack channel.

RishabhMaheshwary commented 3 years ago

Sure, first I will try to understand the implementation of existing attack recipes and then will open a PR for the new recipe.

srujanjoshi commented 2 years ago

Closed since a corresponding Pull Request has been created.