sdsa 1.2.1

Self-hostable data sanitization app with PII detection, pseudonymization, k-anonymity, and bounded DP noise

SDSA

Secure Data Sanitization App (SDSA) is a self-hostable tool for sanitizing tabular data before it leaves a trusted environment.

It ingests CSV, delimited TXT, and single-table SQL INSERT dumps; detects likely sensitive fields; applies explicit per-column privacy policies; enforces k-anonymity; measures or enforces l-diversity for sensitive cleartext attributes; and exports a sanitized CSV with JSON and Markdown privacy reports.

Install

pip install sdsa
sdsa start

Open http://127.0.0.1:8000/.

What It Does

• Serves a browser UI and REST API from one FastAPI application.
• Detects likely PII such as email, phone, card number, government ID, date of birth, name, address, and identifier fields.
• Supports retain, mask, hash, tokenize, redact, drop, numeric_bin, date_truncate, string_truncate, and dp_laplace actions.
• Applies bounded Laplace noise to numeric columns when differential privacy is configured with explicit epsilon, lower, and upper values.
• Tracks cumulative per-column DP epsilon for the uploaded session to prevent repeated noisy releases from being averaged.
• Enforces k-anonymity over selected quasi-identifiers and can enforce l-diversity on sensitive cleartext attributes.
• Provides preflight suppression estimates before processing.
• Runs headlessly with sdsa process for CI/CD and data pipelines.
• Stores uploaded data in memory with a default 30-minute session TTL.

CLI

sdsa start
sdsa start --host 0.0.0.0 --port 8000
sdsa start --random-port
sdsa start --reload

# Batch mode: no server, writes sanitized CSV + JSON/Markdown reports.
sdsa process data.csv --out-dir ./sanitized -k 5
sdsa process data.csv --policy request.json --out-dir ./sanitized

sdsa-server remains an equivalent compatibility alias. The package includes the static frontend, so no separate web build is required.

Privacy Model

SDSA produces pseudonymized microdata with optional per-column local-DP style noise. It does not claim dataset-level (epsilon, delta) differential privacy. Linkage attacks using auxiliary data may still succeed.

k-anonymity bounds prosecutor re-identification risk to at most 1/k for the declared quasi-identifier set, subject to the limits described in each generated privacy report.

l-diversity is measured by default for cleartext non-QI attributes and can be enforced with l >= 2. Homogeneous sensitive groups appear as warnings in the report when l-diversity is measured but not enforced.

Reports also include an information-loss utility score. Downloadable reports strip exact source-side cardinalities, null counts, and numeric bounds while retaining enough metadata to audit field treatment.

Deployment

For production, run sdsa start behind TLS termination and keep one SDSA process per deployment unless you replace the in-memory session store with shared infrastructure. The GitHub repository includes Docker, Compose, nginx, and CI/CD examples.

Links

• Source: https://github.com/defai-digital/sdsa
• Documentation: https://github.com/defai-digital/sdsa/blob/main/README.md
• Deployment guide: https://github.com/defai-digital/sdsa/blob/main/docs/deployment.md
• Privacy model: https://github.com/defai-digital/sdsa/blob/main/docs/privacy-model.md