Oblivious Multi-Party Machine Learning on Trusted Processors
summary
Framework uses real Intel SGX (Skylake) enclaves and a library of data-oblivious primitives plus an oblivious batch shuffling sampler to prevent cache-line side channels. Implementation leverages VP-gather vector loads to hide fine-grained accesses, handles datasets larger than the 94 MB enclave by encrypted batching, and reports low overheads (neural nets and SVMs about 1–7% with SGX, average ~20% vs. plaintext; decision trees can cost much more depending on depth).
AI-written summary · about
chapters (13)
- 0:00Start
- 5:58Protocol Overview
- 6:55Protection Layers
- 7:15Secure Enclaves with Intel SGX
- 8:14Memory Side-channel Protection
- 10:16Library of oblivious primitives libo
- 13:24Supervised Learning
- 14:20Oblivious Data Sampling
- 15:38Oblivious Decision Trees
- 17:56Experimental Setup
- 19:59Are we data oblivious?
- 21:55Trees: Oblivious Code Traces
- 22:24Conclusion
related talks
42mSponsored Session: Confidential Computing - New Capabilities for New Workloads- 42mSponsored Session: Confidential Computing - New Capabilities for New Workloads
28mGenerative AI + Proprietary Data
19mRethinking Isolation Mechanisms for Datacenter Multitenancy
35mVeracruz: Privacy-preserving mutliparty computation using trusted hardware
32mBeyond Software: Hardware-Based Security
01h 00mPrivacy Preserving Smart Contracts
20mConfidential Computing within an AI Accelerator