A method and system for security vulnerability analysis of learning-based indexing and recommendation systems

Abstract

This invention discloses a security vulnerability analysis method and system for learning-based index recommendation systems. The method includes: First, a method for detecting index preferences is proposed under a black-box scenario. Based on the recommendation results of different workloads using the learning-based index recommendation model, the method extracts the preference ranking of different index columns to guide the poisoning attack steps. Second, an attack workload is designed to trap the learning-based index recommendation model in a local optimum. This attack workload is then mixed into the normal working workload and injected into the learning-based index recommendation system to achieve poisoning. This invention can effectively poison existing learning-based index recommendation systems, significantly reducing their index recommendation effectiveness.

Images