Reproduce “Black-box Dataset Ownership Verification via Backdoor Watermarking” on OpenResearch

About this paper

Yiming Li, Mingyan Zhu, Xue Yang, Yong Jiang, Tao Wei, Shu-Tao Xia

Dataset Verification via Backdoor Watermarking (DVBW) enables black-box verification of dataset ownership by embedding hidden backdoor watermarks directly into released datasets, allowing a defender to check if a suspicious model was trained on their data. The method achieves high watermark success rates (e.g., >94% for images, nearly 100% for NLP and graph data) with minimal impact (<2% reduction) on dataset utility across various data modalities.

Key insights

• Backdoor attacks, traditionally adversarial, can be re-purposed for defensive purposes like dataset ownership verification, demonstrating a novel application of adversarial AI techniques.
• The framework addresses a crucial, previously unaddressed problem in AI intellectual property protection, particularly in the practical black-box setting where only model outputs are accessible.
• The method is broadly applicable across diverse data modalities (images, natural language, graphs) and deep learning tasks, highlighting its versatility and generalizability.

Summary by alphaXiv.