GraphReplica leakage testing
Prove your data did not leak
Catch privacy risk before training starts. GraphReplica leakage tests every run and gives governance teams audit-ready evidence that no sensitive value survived into your synthetic data.
See the evidence
Watch leakage drop to zero
Step through real data, ordinary synthetic, and Secludy. Then toggle the category breakdown to see every sensitive type tested and cleared.
Leakage by method
Sensitive values recoveredRaw records, no protection. 100% of injected sensitive values were recovered, so this run should be blocked before training.
Category breakdown
- Canaries injected
- 10,000
- Values recovered
- 2,750
- Result
- Blocked, sensitive values exposed
How testing works
Every run is tested before release
Leakage testing runs on every replica, scores the risk by attack and by category, and confirms the data is still useful before you ship it to AI teams.
Membership inference testing
We attack your replica the way an adversary would and measure whether the output reveals who was in the source data. You learn the real exposure before a model ever sees it.
Canary insertion and extraction checks
We inject unique values into the source, generate the replica, then measure how many survive into the output. Secludy holds the count at zero, so you can prove nothing escaped.
Category-level leakage breakdown
Bitcoin wallets, vehicle IDs, drivers licenses, and SSNs each get their own score. You see exactly where unprotected data fails and confirm Secludy drives every category to zero.
Audit-ready reports
Each run ships with the protection settings, data lineage, and reproducibility metadata. Legal, security, and compliance teams get evidence they can stand behind, built to meet GDPR, CCPA, and HIPAA requirements.
Useful data, proven safe
We benchmark every replica against a real holdout, so zero leakage never means zero value. Utility stays within about 5% of real-data performance, and you get the numbers in the same report.
Block a risky dataset
When a run exposes sensitive values, GraphReplica stops it before training. Risky data never reaches a model, and your teams keep a clear record of why it was held back.
Ship AI data your governance teams approve
See GraphReplica leakage test a dataset live and walk away with an audit-ready report.