Establishing Link...
SecureEncrypted
Demo Active
Live demo — Adraca Synthetic v1.2.0
Anonymization Middleware for Secondary R&D
Adraca Synthetic
Differentially-private synthetic patient data from a Gaussian Copula engine: (ε,δ)-DP synthesis with a per-dataset epsilon budget ledger, DCR/NNDR re-identification risk scoring against a 0.09 threshold with automatic retry, an HMAC-chained HIPAA-style audit log, and JSON-LD provenance certificates. Designed against EHDS requirements. Live on AWS EKS at spe.adraca.io.
v1.2.0 · AWS EKS · live at spe.adraca.io · demo endpoint generates synthetic rows
Interface & Capabilities
Synthesize form — rows, ε and DP mechanism
Completed run — 3.3% re-ID risk, threshold passed
EDPB three-criteria test on the generated cohort
DCR / NNDR distributions with DPO threshold
Synthetic cohort preview with CSV download
Fidelity Lab — KL, Hellinger, KS, Frobenius
Risk-scored synthesis
Every job is scored with DCR and NNDR re-identification metrics over Gower distance; a job failing the 0.09 risk threshold automatically retries at lower ε, up to three times.
Privacy budget enforcement
(ε,δ)-differentially-private Gaussian Copula synthesis with a per-dataset epsilon ledger and spending cap, resettable only through a DPO endpoint.
Tamper-evident audit
HIPAA-style append-only audit log, HMAC-SHA256-chained with an on-demand verification endpoint, plus a GDPR Art. 17 erasure endpoint.
Architecture
Synthesis Engine
SDV Gaussian Copula with (ε,δ)-DP noise; Higham nearest-PSD correction keeps correlation matrices valid.
Temporal Constraints
Enforces temporal ordering constraints on generated clinical events.
Risk Scoring (Gower Distance)
DCR and NNDR metrics computed over mixed categorical and numeric columns to measure re-identification risk.
Threshold Enforcement
Risk above 0.09 triggers automatic re-synthesis with reduced ε, up to three retries, before a job can pass.
Stage 01
Upload & Queue
CSV/Parquet upload (200 MB cap) to S3; synthesis runs as an async job on a persistent queue.
Stage 02
DP Synthesis
Gaussian Copula fitted with (ε,δ)-DP noise; ε debited from the dataset's budget ledger.
Stage 03
Risk Scoring
DCR/NNDR computed against the source; 0.09 threshold enforced with auto-retry at lower ε.
Stage 04
Certify & Deliver
PDF fidelity report and JSON-LD provenance certificate issued; output delivered via presigned S3 URLs.
Deployment Specifications
Algorithm
Gaussian Copula (SDV) · (ε,δ)-DP
Privacy Metric
DCR / NNDR over Gower distance
Risk Constraint
0.09 threshold · auto-retry at lower ε
Compliance
GDPR Art. 17 · HIPAA-style audit · EHDS-aligned design
adraca --log-stream
[--:--:--]UPLOAD RECEIVED [CSV -> S3]...
[--:--:--]SYNTHESIS JOB QUEUED [RQ]...
[--:--:--]FITTING GAUSSIAN COPULA [(ε,δ)-DP]...
[--:--:--]EPSILON LEDGER DEBITED [BUDGET OK]...
[--:--:--]SCORING RISK [DCR/NNDR · GOWER]...
[--:--:--]RISK BELOW 0.09 THRESHOLD [PASS]...
[--:--:--]AUDIT ENTRY CHAINED [HMAC-SHA256]...
[--:--:--]CERTIFICATE ISSUED [JSON-LD + PDF].