This research introduces a novel approach to detect unauthorized use of datasets in Retrieval-Augmented LLMs through watermarked canaries - specially crafted data entries that can prove dataset ownership.

• Achieves 95-99% detection accuracy of unauthorized dataset use
• Creates synthetic, watermarked data that blends naturally with authentic content
• Provides dataset owners with tools to identify IP infringement
• Demonstrates effectiveness across various LLM architectures and retrieval methods

As RAG systems become standard in AI deployments, this technique offers a crucial security layer for organizations concerned about protecting proprietary datasets and intellectual property rights in an increasingly complex AI landscape.

Dataset Protection via Watermarked Canaries in Retrieval-Augmented LLMs