Model provenance
Also called Model Provenance
Knowing exactly where a model came from — who trained it, on what, and that it wasn’t tampered with since.
Think of it like
A used car’s full service history and title — proof of where it’s been and that no one swapped the engine.
Example
Before deploying a downloaded checkpoint, a team verifies its cryptographic hash and signed lineage to confirm it is the untouched original.
How it actually works
Provenance tracks a model’s lineage — base model, training data, fine-tunes — plus integrity signals like checksums and signatures. It is the main defense against supply-chain tampering, since a trojan or backdoor is invisible in behavior but shows up as a broken chain of custody. Emerging practices include signed model cards and attestation.
For product teams
The paper trail that lets you trust a model you didn’t train yourself.
For engineers
Track lineage plus cryptographic integrity (hashes, signatures); it catches supply-chain tampering that behavioral tests miss.
Related
- Supply chain risk — The main defense against supply chain risk.
- Backdoor attack — Helps detect a hidden backdoor attack.
- Model card — Often documented alongside a model card.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome