Crate opendp

source ·
Expand description

A library for working with differential privacy.

This library implements the framework described in the paper, A Programming Framework for OpenDP. OpenDP (the library) is part of the larger OpenDP Project.


OpenDP provides three main concepts:

  • A flexible architecture for modeling privacy-preserving computations.
  • Implementations of several common algorithms for statistical analysis and data manipulation, which can be used out-of-the-box to assemble DP applications.
  • Facilities for extending OpenDP with new algorithms, privacy models, etc.

§User Guide

A more thorough User Guide can be found on the docs website.

OpenDP applications are created by using constructors and combinators to create private computation pipelines. These can be written directly in Rust, or by using a language binding that uses OpenDP through an FFI interface. Python is the first language binding available, but we plan to add others in the future.

§Rust Application Example

Here’s a simple example of using OpenDP from Rust to create a private sum:

use opendp::error::Fallible;

#[cfg(all(feature = "untrusted", feature = "partials"))]
pub fn example() -> Fallible<()> {
    use opendp::transformations::{make_split_lines, then_cast_default, make_cast_default, then_clamp, then_sum};
    use opendp::combinators::{make_chain_tt, make_chain_mt};
    use opendp::measurements::then_laplace;

    let data = "56\n15\n97\n56\n6\n17\n2\n19\n16\n50".to_owned();
    let bounds = (0.0, 100.0);
    let epsilon = 1.0;
    // remove some epsilon to account for floating-point error
    let sigma = (bounds.1 - bounds.0) / (epsilon - 0.0001);

    // Construct a Transformation to parse a csv string.
    let split_lines = make_split_lines()?;

    // The next transformation wants to conform with the output domain and metric from `split_lines`.
    let cast = make_cast_default::<_, String, f64>(

    // Since the domain and metric conforms, these two transformations may be chained.
    let load_numbers = make_chain_tt(&cast, &split_lines)?;
    // You can use the more convenient `>>` notation to chain instead.
    // When you use the `then_` version of the constructor,
    //     the `>>` operator will automatically fill the input domain and metric from the previous transformation.
    let load_and_clamp = load_numbers >> then_clamp(bounds);
    // After chaining, the resulting transformation is wrapped in a `Result`.
    let load_and_sum = (load_and_clamp >> then_sum())?;

    // Construct a Measurement to calculate a noisy sum.
    let noisy_sum = load_and_sum >> then_laplace(sigma, None);

    // The same measurement, written more succinctly:
    let noisy_sum = (
        make_split_lines()? >>
        then_cast_default() >>
        then_clamp(bounds) >>
        then_sum() >>
        then_laplace(sigma, None)

    // Check that the pipeline is (1, 1.0)-close
    assert!(noisy_sum.check(&1, &epsilon)?);

    // Make a 1.0-epsilon-DP release
    let release = noisy_sum.invoke(&data)?;
    println!("release = {}", release);
#[cfg(all(feature = "untrusted", feature = "partials"))]

§Contributor Guide

A more thorough Contributor Guide can be found on the docs website.

§Adding Constructors

Measurement constructors go in the module crate::measurements, Transformation constructors in the module crate::transformations, and Combinator constructors in the module crate::combinators.

There are two code steps to adding a constructor function: Writing the function itself, and adding the FFI wrapper.

§Writing Constructors

Constructors are functions that take some parameters and return a valid Measurement or Transformation. They typically follow a common pattern:

  1. Choose the appropriate input and output Domain.
  2. Write a closure that implements the Function.
  3. Choose the appropriate input and output Metric/Measure.
  4. Write a closure that implements the PrivacyMap/StabilityMap.
§Example Transformation Constructor
pub fn make_i32_identity() -> Transformation<AtomDomain<i32>, AtomDomain<i32>, AbsoluteDistance<i32>, AbsoluteDistance<i32>> {
    let input_domain = AtomDomain::default();
    let output_domain = AtomDomain::default();
    let function = Function::new(|arg: &i32| -> i32 { *arg });
    let input_metric = AbsoluteDistance::default();
    let output_metric = AbsoluteDistance::default();
    let stability_map = StabilityMap::new_from_constant(1);
    Transformation::new(input_domain, output_domain, function, input_metric, output_metric, stability_map).unwrap()
§Input and Output Types

The Function created in a constructor is allowed to have any type for its input and output Domain::Carrier. There’s no need for special data carrying wrappers. The glue code in the FFI layer handles this transparently. However, the most common are the Rust primitives (e.g., i32, f64, etc.), and collections of the primitives (Vec<i32>, HashMap<String, f64>).

§Handling Generics

Measurement/Transformation constructors are allowed to be generic! Typically, this means that the type parameter on the constructor will determine type of the input or output Domain::Carrier (or the generic type within, for instance the i32 of Vec<i32>).


  • Convert between noise scales and accuracies.
  • Various combinator constructors.
  • Core concepts of OpenDP.
  • Framework for flexible abstract data type model for DataFrames.
  • Various implementations of Domain.
  • Error handling utilities.
  • FFI bindings for OpenDP.
  • Various measurement constructors.
  • Various implementations of Measures (and associated Distance).
  • Various implementations of Metrics (and associated Distance).
  • Traits that enable building stable and private algorithms.
  • Various transformation constructors.