SHF: Small: Hardware-Software Co-design for Privacy Protection on Deep Learning-based Recommendation Systems

University Of Pittsburgh, Pittsburgh PA

Investigators

Abstract

Deep-learning recommendation models (DLRMs) are widely adopted in industries as they exploit deep-learning neural networks to provide personalized recommendations, which significantly enhance user experience and customer loyalty. The sizes of large DLRMs are dominated by their embedding tables, which may be of 10s or even 100s of gigabytes, making it an appealing choice to save such tables in cloud servers. Unfortunately, querying the embedding tables on cloud servers demands sending in clients' private data and thus may leak their sensitive information. The privacy issue has become a major challenge when deploying large DLRMs in the cloud. Adopting traditional tree-based oblivious memory protocols can provide highly robust privacy protection, but suffer from significant performance degradation and low memory space utilization. This project focuses on developing a hardware-software co-design framework to address the performance and space issues of privacy protection for cloud-based DLRMs. The project’s outcomes will have significant societal impact by servicing personalized recommendations securely in the clouds. The investigators will actively recruit and train undergraduate and graduate students from underrepresented groups, and provide research and education opportunities for K-12 students. The project centers around three innovative approaches: (1) A hardware-software co-design for significant performance improvement; (2) The effective integration of non-volatile memory for significant memory space utilization improvement; (3) The exploration of distinct DLRM access behaviors for secure protocol designs. In particular, the project exploits the on-chip trusted computing hardware on modern processors and the distinguishing characteristics of DLRMs such that secure operations can be partitioned among clients and trusted processors for effective privacy protection with low overhead. The project contains concrete steps to develop schemes for one security group of users accessing one cloud server as well as for multiple security groups of users accessing multiple cloud servers. Successfully addressing the performance and space issues of privacy protection can advance the modern computing paradigm in the AI era, i.e., how the large DLRMs as well as modern AI models are deployed. 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.