Guardians of the Dataverse: Securing the AI Supply and Data Chain
summary
Maps NIST SP 800-204D thinking and in-toto attestation onto centralized AI/ML pipelines, arguing for signed CI/CD-produced models, data provenance, and differential privacy for sensitive training. Builds and demos a new chunked-hash tool called terapin (Go and R reference impls) that splits data into 2 MiB chunks, yields a 32-byte root, and lets you validate slices — 1 PB would take ~27 hours with naive SHA-256 but terapin turns a 16 GiB file from 9.5s to 2.7s locally.
AI-written summary · about
chapters (7)
- 0:00Securing AI pipelines
- 3:32AI supply chains explained
- 8:08Securing AI pipelines
- 12:03Model lifecycle and data provenance
- 18:47Efficient dataset attestation with Terapin
- 27:20Terapin hashing demo
- 30:33Q&A
related talks
23mTrust, Track, and Verify: Securing AI Pipelines End-to-End
19mWe Know Security but How Do We Secure GenAI End-to-End?
17mPreserving integrity and trust in AI
17mTrust at Scale: Fueling AI and Defending Against Quantum Threats
36mSecurity for AI and Multi-Party Collaboration with Confidential Computing (Oh, and...
19mSecuring AI Pipelines from Development to Production
09mSecuring Generative AI Workloads
39mSecuring AI at Scale: Practical Defenses against Prompt...