Navigating the Performance-Security Trade-Off in Future Analytics on Shared Data

Securing analytics on shared data is important but expensive. Analyzing datasets from multiple data owners can yield valuable insights [1, 2, 3, 4, 5] but poses significant security risks. Even within enterprises - our primary focus - precautions are necessary when handling data across subsidiaries and geographic regions [6, 7]. Existing security solutions based on Trusted Execution Environments (TEEs) [8, 9], fully homomorphic encryption [10], and structured encryption [11] offer strong protections, albeit in a physically centralized manner. For more decentralization, there are exciting approaches based on Secure Multi-Party Computation (MPC) [12] that do not need a trusted third party nor merging datasets at a central location. Recent projects [6, 13, 14, 15] show that MPC can reduce the risk of leaks for analytics on shared data under stronger security guarantees. However, MPC queries are often impractically slow, requiring orders of magnitude more computation and communication than plain-text or TEEbased query execution.