Ready reckoner
Data governance
- Data owner decides whether data may be used for training or retrieval. A provider that trains on your data for its own purposes becomes a controller.
- Residency (where processed) · sovereignty (whose law) · localization (legal duty to keep in-country).
- Classification travels into chunks, embeddings and models; derived artifacts take the highest input classification.
- Provenance = where it came from · lineage = where it went · chain of custody = proof it wasn't changed.
Privacy techniques
| Technique | Reversible? | Still personal data? |
|---|---|---|
| Masking / tokenization / pseudonymization | Yes (with key) | Yes |
| Anonymization | No (if done well; risk-measured) | No |
| Differential privacy | No | Output protected by ε |
- Quasi-identifiers (postcode, DOB, industry) single people out.
- DP: smaller ε = stronger privacy, less accuracy; losses compose → cap the budget.
- Erasure: easy from retrieval; hard from weights (retrain or unlearn). Design for erasure before training.
Training data security
- Protect with access control, separation of duties, validation, hashes, signed manifests verified before training, lineage.
- Poisoning: availability (degrades all) · targeted (chosen inputs) · backdoor (hidden trigger, passes normal tests).
- Privacy attacks: extraction, membership inference, model inversion. Reduced by dedup, DP, regularization, output minimization.
- Synthetic data: validate, label as synthetic, limit recursion.
Protection tooling
| Tool | Covers |
|---|---|
| DSPM | Where data is |
| CSPM | Infrastructure configuration |
| DLP | Data leaving (pattern + semantic for AI) |
| AI-SPM | AI assets, shadow AI |
- Keys in KMS/HSM/vault, never in notebooks or prompts; rotate on exposure. Crypto-erase works only if every copy was encrypted and every key copy is destroyed.
- Sanitization: clear · purge · destroy.
Model lifecycle and CIA
- Model owner + business owner; evidence scales with tier. Registry is a control only if the pipeline deploys nothing else.
- Pickle executes code on load → safetensors, scan, isolate. Hash and sign; verify at load.
- Model output is untrusted input to whatever consumes it.
- Release: shadow → canary → automatic rollback on guardrail metrics.
- Extraction uses the API: watch volume, near-duplicates, coverage; minimize outputs; quotas; contract ban on training on outputs.
- Watermark = deliberate mark (per licensee) · fingerprint = measured existing traits.
- Availability = cost: token-aware limits, budget caps, loop caps, graceful degradation.
Supply chain
- Assess external datasets (poisoning, rights, PII, bias, provenance). Pin by hash, internal mirrors.
- Watch typosquatting, namespace reuse, dependency confusion, compromised releases.
- After a compromised release: rebuild from verified images, rotate every exposed secret.
- Foundation-model risk is inherited: evaluate for your use, re-test on every change, plan exit.