Fine Grained Access Control in Vector Databases

Thu, 20 Feb 2025 00:00:00 +0000

Vector databases are particularly well-suited for similarity search using search algorithms like approximate nearest neighbor (ANN) search and they are used in development of Retrieval-Augmented Generation (RAG) systems, to reduce hallucinations in responses of AI systems. One significant challenge in using vector databases, especially in applications like RAG, is ensuring data privacy and security. For example, a clothing company that builds an AI chatbot that interacts with a vector database containing customer orders and product data could expose sensitive customer information without proper access restrictions. Incorporating Fine-Grained Access Control in vector databases is important for enforcing user preferences on data sharing and complying with privacy regulations. This project explores how to embed fine-grained access control within vector databases to ensure secure and privacy-compliant query answering.

Tattletale

Sat, 01 Jun 2024 00:00:00 +0000

Tattletale uses denial constraints to discovery data inferences inside of a database relative to sensitive cells. The cells that make up the denial constraints are then checked to see which cells infer information on them. In the end all the cells that infer data on the sensitive cells and the cells that could be used to reconstruct those inferences are placed into a list which is used to generate a view that does not contain those cells. Since inference can only be reconstructed as long as only one predicate is missing we can use that to minimize how many cells we have to hide. The benefit of Tattletale is that it provides protection against inference which access control lists don’t provide. The current challenge is trying to improve the run time performance and decrease the number of cells that have to be hidden while also guaranteeing a certain level of protection against reconstruction.Tattletale uses denial constraints to discovery data inferences inside of a database relative to sensitive cells. The cells that make up the denial constraints are then checked to see which cells infer information on them. In the end all the cells that infer data on the sensitive cells and the cells that could be used to reconstruct those inferences are placed into a list which is used to generate a view that does not contain those cells. Since inference can only be reconstructed as long as only one predicate is missing we can use that to minimize how many cells we have to hide. The benefit of Tattletale is that it provides protection against inference which access control lists don’t provide. The current challenge is trying to improve the run time performance and decrease the number of cells that have to be hidden while also guaranteeing a certain level of protection against reconstruction.

Inference Control | DIPr Lab at PSU

Fine Grained Access Control in Vector Databases

Tattletale