This repository serves as a comprehensive, curated list of academic resources on backdoor learning, a critical area of machine learning security. It targets researchers, engineers, and students interested in understanding, developing, or defending against backdoor attacks (also known as neural Trojans) in AI systems. The primary benefit is a centralized, categorized collection of papers, toolboxes, and theses, facilitating efficient research and development in this domain.

How It Works

The repository organizes a vast collection of academic papers, dissertations, and toolboxes related to backdoor learning. Resources are categorized by attack type (e.g., poisoning-based, weights-oriented), defense strategies (e.g., preprocessing, empirical defense, certified defense), and application domains (e.g., federated learning, NLP, computer vision). This structured approach allows users to quickly navigate and find relevant research on specific aspects of backdoor attacks and defenses.

Quick Start & Requirements

• Access: No installation required; it's a curated list of links.
• Requirements: Access to academic paper repositories (e.g., arXiv, conference proceedings) for full content.
• Resources: Links to the survey paper and the GitHub repository itself are provided for deeper dives.

Highlighted Details

• Extensive categorization covering various attack vectors, defense mechanisms, and application domains.
• Includes links to relevant toolboxes like BackdoorBox, TrojanZoo, and OpenBackdoor.
• Features a dedicated section for dissertations and theses, offering in-depth research perspectives.
• Regularly updated with recent conference papers (e.g., ICLR, NeurIPS, CVPR).

Maintenance & Community

The repository is maintained by THUYimingLi, with updates noted for ICLR'23, NeurIPS'22, and AAAI'23 papers. The author expresses a commitment to monthly maintenance, though personal circumstances have caused temporary suspensions. Contributions are welcomed via direct contact or pull requests.

Licensing & Compatibility

The repository itself is not software; it's a collection of links. The licensing of the linked papers and toolboxes would vary by their respective sources. Compatibility for commercial use or closed-source linking depends entirely on the licenses of the individual resources cited.

Limitations & Caveats

The repository is a curated list and does not provide executable code or direct access to all papers; users may need institutional access for some publications. The organization, while comprehensive, relies on the categorization criteria defined by the maintainer.