Downloads n/athis week

GitHub stars 3

Extension repository on GitHub

Extension descriptor (YAML)

Automatic query privatization using the PAC Privacy framework

Installing and Loading

INSTALL pac FROM community;
LOAD pac;

Example

-- Declare a privacy unit table with protected columns
CREATE PU TABLE employees (
    id INTEGER, department VARCHAR, salary DECIMAL(10,2),
    PAC_KEY (id),
    PROTECTED (salary)
);

INSERT INTO employees VALUES
    (1, 'Engineering', 95000), (2, 'Engineering', 110000),
    (3, 'Sales', 80000),      (4, 'Sales', 72000),
    (5, 'Marketing', 85000),  (6, 'Marketing', 90000);

-- Aggregates on protected columns are automatically noised
SELECT department, AVG(salary)::INTEGER AS avg_salary, COUNT(*) AS headcount
FROM employees GROUP BY department;
-- ┌─────────────┬────────────┬───────────┐
-- │ department  │ avg_salary │ headcount │
-- │   varchar   │   int32    │   int64   │
-- ├─────────────┼────────────┼───────────┤
-- │ Marketing   │      95601 │        15 │
-- │ Engineering │     109039 │        16 │
-- │ Sales       │      81145 │        17 │
-- └─────────────┴────────────┴───────────┘

About pac

PAC (Privacy-preserving Automatic Queries) is a DuckDB extension that transparently adds noise to aggregate query results to protect against Membership Inference Attacks. Unlike Differential Privacy, PAC requires no per-query privacy analysis — just declare which columns are protected and queries are privatized automatically.

Key Features:

• CREATE PU TABLE syntax to declare privacy units and protected columns
• PAC_LINK for privacy propagation through join chains
• Supports COUNT, SUM, AVG, MIN, MAX, COUNT(DISTINCT)
• Tunable privacy-utility tradeoff via pac_mi parameter

How It Works:

1. Mark tables with CREATE PU TABLE, specifying a PAC_KEY (privacy unit) and PROTECTED columns
2. Protected columns cannot be projected directly — they can only appear inside aggregates
3. The extension automatically injects calibrated noise into aggregate results
4. Join chains propagate privacy constraints via PAC_LINK foreign key declarations

Configuration Settings:

• pac_mi: Mutual information bound controlling the privacy-utility tradeoff (default: 1/128)

Requirements:

• DuckDB 1.5.0+

Added Functions

Overloaded Functions

| function_name | function_type | description | comment | examples | |—————|—————|————-|———|———-|

Added Types

| type_name | type_size | logical_type | type_category | internal | |———–|———-:|————–|—————|———-|

Added Settings