Building Trust in ML: Mapping the Model Lifecycle for ML Integrity and Transparency
summary
Atlas maps end-to-end ML provenance by linking dataset, training script, model and evaluation artifacts with cryptographic hashes and storing signed CTPA manifests in a transparency backend, demonstrated with Hugging Face MNIST downloads, mocked TDX attestations and a MongoDB backend; the prototype v0.1 integrates with sigstore-style signing, verifies lineage chains and fails when links are tampered, and plans SLSA alignment and GPU/accelerator support. Atlas CTPA manifests Intel TDX
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:01ML lifecycle transparency need
- 3:08ML pipeline supply chain risks
- 5:16Model integrity approaches
- 8:14Atlas for transparency
- 9:57Atlas captures and excludes
- 11:01Atlas architecture and blocks
- 13:47Atlas provenance generation
- 15:14Demo setup and pipeline
- 17:02Generating manifests for data
- 18:48Linking provenance artifacts
- 20:13Validating provenance and tampering
- 21:39Provenance graph visualization
- 22:01Atlas future steps
- 23:46Key takeaways and open source
- 24:58Q&A: Integration time
- 26:24Q&A: Security of unsigned weights
- 27:34Q&A: CTPA vs SLSA provenance
- 29:02Q&A: Transparency logs role
- 30:39Q&A: User-friendly verification
- 31:00Closing remarks
related talks
25mFrom Model To Trust: Building Upon Tamper-proof ML Metadata Records
23mTrust, Track, and Verify: Securing AI Pipelines End-to-End
19mModel transparency for AI/ML security
21mML Model Signing: Cryptographically Paving the Way to Provenance in Machine Learning Models
23mEnd-to-End Secure ML Development
21mEnd-to-end secure ML Development
26mTaming ML’s Wild West: Practical Model Signing with Sigstore on Kaggle
19mWe Know Security but How Do We Secure GenAI End-to-End?