stage S94h
Model Extraction & IP Protection
Protect hosted model endpoints against systematic distillation sweeps and logit harvesting using per-key token/query budgets, logprob stripping, and query-distribution anomaly alerts.
Stage detail
When: After Week 8, only if you operate a hosted model/API or protect a valuable fine-tuned model.
- Rate limits by account, key, IP, tenant, and usage budget; anomaly detection for systematic query patterns.
- Abuse detection for repeated probing, low-temperature systematic queries, and near-duplicate prompt sweeps.
- Understand membership inference and model inversion conceptually, including limits.
- Understand output watermarking's limitations; it does not replace access control or abuse monitoring.
- Protect model endpoints with quotas, monitoring, and appropriate access controls.
Extra practice: simulate extraction-style query patterns against your own mock endpoint and build rate-limit/anomaly alerts. Do not attempt to extract third-party models.
Do not learn yet: research-grade defenses unless model/IP protection is part of the target role.