Collaborative Research: CISE-MSI: RDP: SaTC: A Database Architecture for Enhanced Privacy in Machine-Learning Applications

The evolution of database architectures has prioritized speed, capacity, scalability, and flexibility over privacy considerations, creating challenges in protecting sensitive data like healthcare, biometrics, financial, and educational records. Privacy policies and regulations often fail to integrate seamlessly with data-management practices, complicating the task of ensuring privacy across end-to-end workflows. In addition, decoupling decentralized data management and decentralized machine learning (ML) over federated data silos (i.e., federated learning) causes difficulties in detecting malicious clients and handling diverse data distributions. This project aims to address these shortcomings by developing a database architecture to integrate the privacy regulation and compliance process, enhance federated learning with decentralized data-management functions such as data synthesis and profiling, and automate privacy-model configuration in artificial intelligence (AI) workflows. The project will be evaluated using real-world healthcare and biometrics workloads to demonstrate policy consistency, better privacy-utility tradeoffs, and improvement of productivity in configuring privacy-preserving mechanisms. This project will leverage the partnership between Texas A&M University - Central Texas and Arizona State University to train underrepresented students in both universities at the intersection of AI/ML, privacy, and database systems. The project aims to provide a data-architecture design that enables unified privacy policies, enhance federated learning with advanced data management, and optimize privacy-aware query and storage mechanisms. This project consists of the following research thrusts: (i) developing a database system that integrates AI/ML capabilities, facilitating the coordination of data-privacy policies, AI/ML workflows, and regulatory compliance as well as ensuring that data management and privacy regulations are seamlessly aligned with AI/ML processes; (ii) building a federated data-management framework for federated learning to enhance incentive mechanisms, detect malicious gradients, and balance non-independent and identical distribution to improve the accuracy and robustness of federated learning; and (iii) creating query and storage optimizers that automatically select the appropriate model architecture for privacy-preserving training requests, as well as the storage scheme for the resulting model, thereby ensuring that user utility and privacy objectives are met and enhancing both the effectiveness and the privacy of end-to-end AI/ML workflows. This project will also include educational activities targeting students from underrepresented populations in both universities, including, but not limited to, an annual AI Day activity, students exchange between the two partner universities, and curriculum innovation leveraging the research products of this project. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.