Enabling Low-Cost Secure Computing on Untrusted In-Memory Architectures

Explore a 20-minute conference presentation that addresses the critical security challenges of Processing-in-Memory (PIM) architectures in modern computing systems. Learn how researchers from UC Riverside tackle the fundamental problem of the memory wall—where processor performance is constrained by memory bandwidth—while maintaining data security when computation moves closer to memory storage. Discover how multi-party computation (MPC) techniques, including arithmetic secret sharing and Yao's garbled circuits, enable secure outsourcing of bandwidth-intensive computations to PIM modules without exposing unencrypted data to potential attackers. Understand the implementation of precomputation optimization strategies that prevent CPU bottlenecks during MPC operations, and examine practical applications including Deep Learning Recommendation Model inference and Logistic Regression. Review comprehensive evaluation results using the UPMEM PIM system that demonstrate up to 14.66x performance improvements compared to secure CPU configurations while preserving data confidentiality and integrity for both linear and nonlinear computations.