AI/ML model signing

Threshold-sign ML model manifests to prevent unauthorized model swaps in production. Attracts the responsible-AI community — no other threshold framework covers this use case.

How it works

  1. Model governance quorum (ML engineer, security reviewer, AI safety lead) authenticates via OIDC or share-based threshold.
  2. Model weights are hashed (SHA-256 of safetensors/pickle/ONNX).
  3. Threshold quorum signs the manifest (model name, version, weights hash, safety eval results).
  4. Inference server verifies the manifest before loading.
  5. Signing ceremony anchored to transparency log for AI provenance.

Quickstart

kg = Confium::TC::Cmp20.keygen(3, 5)  # 3-of-5 governance quorum
sig = Confium::TC::Cmp20.sign(kg["shares"].first(3), 3, model_manifest)
# Inference server verifies sig under joint public key before loading.

See examples/ai_model_signing.rb.