Generative AI + Proprietary Data
summary
Shows a practical stack that keeps prompts, RAG DBs and model weights encrypted while running by combining memory encryption and remote attestation on enclave hardware (Intel SGX/TDX, AMD SNP, ARM, Nvidia H100 GPU enclaves). Demonstrates two products: opaque workspace for joint analytics where two silos each encrypt with their own key and an enclave trains/aggregates without revealing raw data (customer: Accenture), and a gateway that strips PII with tokenization while preserving model accuracy. Claims inference overhead ≈5% and argues FHE is orders of magnitude slower.
AI-written summary · about
chapters (8)
- 0:00Introduction to secure AI
- 1:28Security risks in GenAI pipelines
- 3:37How confidential computing works
- 9:16Remote attestation and confidential VMs
- 12:46Using enclaves for GenAI
- 15:56Opaque Systems products overview
- 17:41Protecting prompts with op Gateway
- 19:28Conclusion and Q&A
related talks
37mConfidential AI: Enabling Trustworthy Ecosystems in the Age of Agentic Intelligence - Session
19mHow Can We Have Private Access to Intelligence?
12mKeynote: Protecting Data in the Agentic AI Era
20mKeynote: Future-Proofing Confidential AI
46mOpen for AI: Secure paths to data collaboration volume and diversity
24mStrictly Confidential: Use Cases & Demos of AI Workflows in TEEs
44mGPU-less, Trust-less, Limit-less: Reimagining the Confidential AI Cloud
20mConfidential Computing within an AI Accelerator