DiffPriv is a differential privacy framework for real time data streaming written in Rust. Supporting k-anonymity, (c,l)-diversity and ε-differential privacy. The framework is based on the Preserving Differential Privacy and Utility of Non-stationary Data Streams paper, with various improvements implemented.
This project is the result of my master thesis: Differential privacy in large scale data streaming. It has been developer during an internship at STRM Privacy
it's recommended to first build the application using as it will significantly speed up the algorithm
cargo build --release
An application.conf needs to be present in the root folder. This will build a binary that can be run with the following command
RUST_LOG="debug" ./target/release/diff-priv
This will use a dataset from the datasets folder, the supported datasets can be seen in test/tests.rs
RUST_LOG part can be removed to the users liking. This removes debugging logging when the algorithm will run.
When main.rs is run, the processed datasets can be seen in the exports directory.
Inside the application.conf all the different privacy parameters can be edited to the users liking.
At this moment for buffer_size we use 3*k and for k_max we use 4*k. This can be edited in the environment.rs and tests.rs file.
Additional parameters can be easily added through the config.rs file by adding it as a struct attribute and then adding it to application.conf.
| Parameter | Description |
|---|---|
| k | The minimum k-anonymity level |
| k_max | The maximum k-anonymity level a cluster can have before being deleted |
| l | The minimum l-diversity level |
| (c,l) | The minimum (c,l)-diversity level |
| eps | ε-differential privacy level |
| diff_thrs | The maximum distance the data tuple and the cluster centroid can have |
| delta | The maximum time in seconds a cluster can exist without the addition of new data tuples |
| buff_size | The maximum amount of tuples the buffers W_curr and W_prev can contain |
| noise_thr | categorical noise level |
DiffPriv is a differential privacy framework for real time data streaming written in Rust. Supporting k-anonymity, (c,l)-diversity and ε-differential privacy. The framework is based on the Preserving Differential Privacy and Utility of Non-stationary Data Streams paper, with various improvements implemented.
This library is the result of my master thesis: Differential privacy in large scale data streaming. It has been developer during an internship at STRM Privacy
An example of using the anonymizer can be seen below
use csv::Reader;
use diff_priv::anonymization::microagg_anonymizer::MicroaggAnonymizer;
use diff_priv::noise::laplace::laplace_noiser::LaplaceNoiser;
use diff_priv::test::adult::Adult;
use diff_priv::test::dummy_publisher::DummyPublisher;
// we initialize our noiser that implements the `Noiser` trait
let noiser = LaplaceNoiser::new(0.1, 3, 0.1);
// we initialize a publisher that implements the `Publisher` trait
let publisher = DummyPublisher::default();
// we create the anonymizer with the desired parameters
// k: 2 | k_max: 10 | c: 2 | l: 7 | diff_thres: 0.1 | delta: 10 | buff_size: 5
let mut anonymizer: MicroaggAnonymizer<LaplaceNoiser, Adult, DummyPublisher> =
MicroaggAnonymizer::new(2, 10, 2, 7, 0.1, 10, 5, publisher, noiser);
// load CSV file representing an Adult
let mut file = Reader::from_path("datasets/Adult_1_numeric_only_class_50K.csv").unwrap();
for line in file.deserialize() {
let row_result = line;
// when we call for `anonymizer()` the anonymizer will
// automatically publish to the given backend when the given
// privacy parameter conditions are met
match row_result {
Ok(row) => anonymizer.anonymize(row),
Err(e) => panic!("{}", e)
}
}
// publish remaining data tuples to the given publisher
// in this case a `DummyPublisher`
anonymizer
.cluster_set
.into_iter()
.for_each(|(_, mut cluster)| {
cluster.publish_all(&mut anonymizer.publisher, &mut anonymizer.analysers)
});By implementing the Anonymizable trait on any type of datastructure, DiffPriv will know how to anonymize it.
The following QIs types are implemented
/// value, min_value, max_value, weight of attribute
pub type IntervalType = (
QuasiIdentifierType,
QuasiIdentifierType,
QuasiIdentifierType,
usize,
);
/// rank, max_rank, weight of attribute
pub type OrdinalType = (i32, i32, usize);
/// value, max value, weight of attribute
pub type NominalType = (i32, i32, usize);An example implementation of the Anonymizable trait can be seen below
use std::time::{SystemTime, UNIX_EPOCH};
use serde::{Serialize, Deserialize};
use bimap::BiMap;
use lazy_static::lazy_static;
use uuid::Uuid;
use diff_priv::data_manipulation::anonymizable::{
Anonymizable, QuasiIdentifierType, QuasiIdentifierTypes, SensitiveAttribute,
};
lazy_static! {
static ref CLASS_BIMAP: BiMap<&'static str, i32> =
BiMap::from_iter(vec![("<=50K", 0), (">50K", 1),]);
}
// This is the datastructure that we are going to anonymize
#[derive(Debug, Serialize, Clone, Deserialize)]
pub struct Adult {
timestamp: i32,
age: i32,
capital_gain: i32,
capital_loss: i32,
class: String,
#[serde(skip_deserializing, default = "default_time")]
time_generated: SystemTime,
}
fn default_time() -> SystemTime {
SystemTime::now()
}
impl Default for Adult {
fn default() -> Self {
Self {
timestamp: 0,
age: 0,
capital_gain: 0,
capital_loss: 0,
class: "".to_string(),
time_generated: SystemTime::now(),
}
}
}
impl Adult {
// here we extract an Interval QI from the `age` attribute
fn get_age_qi(&self) -> QuasiIdentifierTypes {
QuasiIdentifierTypes::Interval((
QuasiIdentifierType::Integer(self.age),
QuasiIdentifierType::Integer(1),
QuasiIdentifierType::Integer(100),
1,
))
}
// here we extract an Interval QI from the `capital_gain` attribute
fn get_capital_gain_qi(&self) -> QuasiIdentifierTypes {
QuasiIdentifierTypes::Interval((
QuasiIdentifierType::Integer(self.capital_gain),
QuasiIdentifierType::Integer(0),
QuasiIdentifierType::Integer(100000),
1,
))
}
// here we extract an Interval QI from the `capital_loss` attribute
fn get_capital_loss_qi(&self) -> QuasiIdentifierTypes {
QuasiIdentifierTypes::Interval((
QuasiIdentifierType::Integer(self.capital_loss),
QuasiIdentifierType::Integer(0),
QuasiIdentifierType::Integer(5000),
1,
))
}
}
// Here we implement the `Anonymizable` trait
impl Anonymizable for Adult {
// We extract the QIs from the datastructure and return a `vec` of QIs
fn quasi_identifiers(&self) -> Vec<QuasiIdentifierTypes> {
let age = self.get_age_qi();
let capital_gain = self.get_capital_gain_qi();
let capital_loss = self.get_capital_loss_qi();
vec![
age,
capital_gain,
capital_loss,
]
}
// We update the datastructures QIs with a `vec` of QIs. The `vec` needs to be
// popped in the same order that the QIs are extracted with the `quasi_identifiers`
// function
fn update_quasi_identifiers(&self, mut qi: Vec<QuasiIdentifierTypes>) -> Self {
if let (
QuasiIdentifierType::Integer(capital_loss),
QuasiIdentifierType::Integer(capital_gain),
QuasiIdentifierType::Integer(age),
) = (
qi.pop().unwrap().extract_value(),
qi.pop().unwrap().extract_value(),
qi.pop().unwrap().extract_value(),
) {
Self {
timestamp: self.timestamp,
age,
capital_gain,
capital_loss,
class: self.class.to_owned(),
time_generated: self.time_generated,
}
} else {
panic!("Couldn't Adult with QI's")
}
}
// We extract the sensative attribute from the datastructure
fn sensitive_value(&self) -> SensitiveAttribute {
SensitiveAttribute::String(self.class.to_owned())
}
// We return a vector of strings containing the String version of the QIs
// Used for printing to CSVs
fn extract_string_values(&self, uuid: Uuid, dr: f64) -> Vec<String> {
vec![
uuid.to_string(),
dr.to_string(),
self.timestamp.to_string(),
self.age.to_string(),
self.capital_gain.to_string(),
self.capital_loss.to_string(),
self.class.to_owned(),
]
}
fn get_timestamp(&self) -> SystemTime {
self.time_generated
}
}To publish an anonymized struct to a desired backend we use the Publisher trait.
DiffPriv also support exporting to an Apache Kafka topic. This can be seen in publishing directory.
An example publisher for CSVs can be seen here: CsvPublisher.
To implement a custom publishing backend one can use the Publisher trait.
DiffPriv support Laplace noise for ε-differential privacy. The noiser supports 2 different kind of noise: one for numerical values and one for categorical. To implement a custom implementation of ε-differential privacy noise, one can use the Noiser trait.
The architecture of the DiffPriv framework can be seen below
In my thesis is described tests using knn-test.sh. To run this you need Java 8.
License:
MIT License
Copyright (c) 2022 Maciek Mika
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.