Osprey: Transparent and Efficient Virtual Memory for Secure Computation
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
Osprey provides a runtime virtual memory layer for secure computation that pages encrypted data to SSDs by forking a speculative process and a programmed process, using userfaultfd, an eBPF fallback collector and MPK to avoid kernel changes. It leverages content-oblivious aliasing to map all ciphertext to one physical page and skips crypto in speculation, requiring <200 LOC per SC library and no app changes. Evaluation on garbled circuits and CKKS shows Osprey matches or beats mage in 7/8 workloads and hits >60% of unbounded performance in 6/8, scaling linearly with threads.
AI-written summary · about
chapters (11)
- 0:00Osprey introduction
- 1:02Secure computation challenges
- 2:48Swapping overhead and performance
- 4:31Input obliviousness
- 6:15Efficient virtual memory
- 7:18Content obliviousness and optimization
- 9:45Memory access patterns
- 11:32Process and memory management
- 12:58Evaluation setup
- 13:40Microbenchmark results
- 14:21Real-world performance
related talks
24mOscar: A Practical Page-Permissions-Based Scheme...
19mCoSMIX: A Compiler-based System for Secure Memory Instrumentation and Execution in Enclaves
22mHodor: Intra-Process Isolation for High-Throughput Data Plane Libraries
21mWhen Oblivious is Not: Attacks against OPAM
24mShielding Software From Privileged Side-Channel Attacks
18mHardware-Software Codesign for Mitigating Spectre
15mORC: Increasing Cloud Memory Density via Object Reuse with Capabilities
41mProtected Execution Facility