This repository serves as a curated collection of essential research papers on textual adversarial attacks and defenses (TAAD). It aims to provide a comprehensive and organized resource for researchers and practitioners in Natural Language Processing (NLP) focusing on model robustness and adversarial techniques.

How It Works

The repository categorizes papers into logical sections, including toolkits, survey papers, attack methodologies (classified by perturbation level: sentence, word, char, multi-level), defense strategies, certified robustness, and benchmarks. Each entry typically includes a link to the paper's PDF, and often to associated codebases or demos, facilitating easy access and reproducibility.

Quick Start & Requirements

This is a curated list of papers, not a software toolkit. No installation or specific requirements are needed beyond a web browser to access the linked resources.

Highlighted Details

• Comprehensive categorization of TAAD research, from foundational surveys to specific attack and defense techniques.
• Inclusion of links to codebases and demos for many papers, enabling practical exploration.
• Covers a wide range of attack types, including white-box, black-box, and blind attacks, across various granularities.
• Features papers on certified robustness and benchmarking, crucial for evaluating and ensuring model reliability.

Maintenance & Community

The list is maintained by Chenghao Yang at UChicago, with previous contributions from Fanchao Qi and Yuan Zang at THUNLP. Contributions are welcomed.

Licensing & Compatibility

This repository is a collection of links to external research papers. The licensing and compatibility of individual papers are determined by their respective publishers and authors.

Limitations & Caveats

As a curated list, the repository's content is limited to the papers that have been identified and added by the maintainers. It does not include any executable code or direct implementation of the discussed techniques.