Deployments Registry
Inspired by Differential Privacy in Practice: Expose your Epsilons! by Cynthia Dwork, Nitin Kohli, and Deirdre Mulligan, this registry provides:
A publicly available communal body of knowledge about differential privacy implementations that can be used by various stakeholders to drive the identification and adoption of judicious differentially private implementations -Dwork Kohli Mulligan 2019
Deployments Registry
| Tier | Product | Description | Year | Flavor name | Privacy Loss | Model | Accounting | Implementation |
|---|---|---|---|---|---|---|---|---|
|
|
Popular Emojis
by Apple
|
Summary statistics of frequency histograms of emoji usage per locale.
Summary statistics of frequency histograms of emoji usage per locale.
|
2017 | Pure DP |
Event-level
ε: 4.0 |
Local | - | - |
|
|
Assistive AI
by Microsoft
|
"Currently we use differentially private algorithms with privacy parameters (\\(\epsilon\\)=4, \\(\delta\\)< 10-7)."
"Currently we use differentially private algorithms with privacy parameters (\\(\epsilon\\)=4, \\(\delta\\)< 10-7)."
|
2020 | Approximate DP |
User-level
ε: 4.0δ: 1.0e-07 |
Central | - | - |
|
|
Audience Engagement API
by LinkedIn
|
"With these proposed parameter values (...) and applying Theorem 2, we get a final (34.9, 7 × 10−9)-DP monthly guarantee."
"With these proposed parameter values (...) and applying Theorem 2, we get a final (34.9, 7 × 10−9)-DP monthly guarantee."
|
2020 | Approximate DP |
User-level
ε: 34.9δ: 7.0e-09 |
Central | - | - |
|
|
AutoPlay Intent
by Apple
|
"For this use case, we set the parameters for CMS to be m = 1024, k = 65,536, and \\(\epsilon\\) = 8."
"For this use case, we set the parameters for CMS to be m = 1024, k = 65,536, and \\(\epsilon\\) = 8."
|
2017 | Pure DP |
User-level
ε: 8.0 |
Local | - | - |
|
|
Israel's National Registry of Live Births of 2014
by Israeli Ministry of Health
|
2024 |
Birth-level
ε: 9.98 |
Central | - | - | ||
|
|
Broadband Coverage
by Microsoft
|
Publicly available U.S. Broadband Coverage dataset that reports broadband coverage percentages at a zip code-level.
Publicly available U.S. Broadband Coverage dataset that reports broadband coverage percentages at a zip code-level.
|
2021 | Pure DP |
User-level
ε: 0.2 |
Central | - | - |
|
|
Disclosure Avoidance System for Redistricting Data
by U.S. Census Bureau
|
Set of summary statistics (tables with counts of individuals, households, group quarters residents, or housing units with certain characteristics) published as the 2020 Census Redistricting Data (P.L. 94-171) Summary File.
Set of summary statistics (tables with counts of individuals, households, group quarters residents, or housing units with certain characteristics) published as the 2020 Census Redistricting Data (P.L. 94-171) Summary File.
|
2021 | Zero-concentrated DP |
Person-level; housing-unit-level
ρ: 2.63 |
Central | - | - |
|
|
Census County Business Patterns
by United States Census Bureau
|
"Noise added in the PRDP stage is randomly sampled from a discrete Gaussian distribution, and the total rho-zCDP privacy loss budget is 12.058."
"Noise added in the PRDP stage is randomly sampled from a discrete Gaussian distribution, and the total rho-zCDP privacy loss budget is 12.058."
|
2023 | Zero-concentrated DP |
Business establishment
ρ: 12.058 |
Central | - | - |
|
|
COVID-19 Exposure Notification Framework
by Google, Apple
|
"The central differential privacy guarantees \\(\epsilon_c\\) that result from unlinkable aggregation of \\(\epsilon_0\\) = 8 metric report binary vectors, as used in ENPA."
"The central differential privacy guarantees \\(\epsilon_c\\) that result from unlinkable aggregation of \\(\epsilon_0\\) = 8 metric report binary vectors, as used in ENPA."
|
2021 |
User-level
ε: 8.0δ: 1.0e-06 |
Varies | - | - | |
|
|
COVID-19 Search Trends Symptoms
by Google
|
Summary statistics, published as an aggregated and anonymized dataset. It provides daily or weekly time series for various geographic regions, showing the relative frequency of Google searches for approximately 400 predefined symptoms. These values are normalized by the total search activity in the corresponding region to show trends over time.
Summary statistics, published as an aggregated and anonymized dataset. It provides daily or weekly time series for various geographic regions, showing the relative frequency of Google searches for approximately 400 predefined symptoms. These values are normalized by the total search activity in the corresponding region to show trends over time.
|
2020 | Pure DP |
User-day
ε: 1.68 |
Central | - | - |
|
|
Community Mobility Place Visits
by Google
|
Summary statistics, published as the COVID-19 Community Mobility Reports, that charts percentage changes in movement trends for various location categories relative to a historical baseline. These trends analyze the number of visits to public places like retail, parks, and workplaces, and also measure the change in duration of time spent in residences.
Summary statistics, published as the COVID-19 Community Mobility Reports, that charts percentage changes in movement trends for various location categories relative to a historical baseline. These trends analyze the number of visits to public places like retail, parks, and workplaces, and also measure the change in duration of time spent in residences.
|
2020 | Pure DP |
User-day
ε: 2.64 |
Central | - | - |
|
|
HealthKit
by Apple
|
"We deploy CMS to observe the most popular health data types with the following parameters: m = 256, k = 65,536, and \\(\epsilon\\) = 2."
"We deploy CMS to observe the most popular health data types with the following parameters: m = 256, k = 65,536, and \\(\epsilon\\) = 2."
|
2017 | Pure DP |
User-level
ε: 2.0 |
Local | - | - |
|
|
Korean Government Statistical Data Hub Platform
by KOSTAT (National Statistics Office of Korea)
|
"Statistics Korea is promoting the establishment of a public
big data system that leverages cutting-edge privacy-preserving techniques to enable
the safe linkage and use of scattered governmental data."
"Statistics Korea is promoting the establishment of a public
big data system that leverages cutting-edge privacy-preserving techniques to enable
the safe linkage and use of scattered governmental data."
|
2021 |
Aggregate-level
|
Central | - | - | |
|
|
LinkedIn Hiring Reports
by LinkedIn
|
"Therefore each report is (1.2, 10−10)-DP because the ratio
that uses both counts accumulates the privacy loss. Therefore, for each report
date, we guarantee the top-employer reports are together (4.8, 4 × 10−10)-DP."
"Therefore each report is (1.2, 10−10)-DP because the ratio
that uses both counts accumulates the privacy loss. Therefore, for each report
date, we guarantee the top-employer reports are together (4.8, 4 × 10−10)-DP."
|
2020 | Approximate DP |
Event-Month-level
ε: 4.8δ: 4.0e-10 |
Central | - | - |
|
|
Lookup Hints
by Apple
|
"The Apple differential privacy implementation incorporates the concept of a per-
donation privacy budget (quantified by the parameter epsilon), and sets a strict limit on
the number of contributions from a user in order to preserve their privacy... For Lookup Hints, Apple uses a privacy budget with epsilon
of 4, and limits user contributions to two donations per day."
"The Apple differential privacy implementation incorporates the concept of a per-
donation privacy budget (quantified by the parameter epsilon), and sets a strict limit on
the number of contributions from a user in order to preserve their privacy... For Lookup Hints, Apple uses a privacy budget with epsilon
of 4, and limits user contributions to two donations per day."
|
2016 | Pure DP |
User-level
ε: 8.0 |
Local | - | - |
|
|
Mobility Trends During Hurricane
by Spectus
|
"In this case, the Laplace mechanism is used for all four
operations. Finally, the privacy budget, E, is set as follows: • E = 4 for CTU
• E = 2 for CE • E = 1 for CED • E = 3 for SD."
"In this case, the Laplace mechanism is used for all four
operations. Finally, the privacy budget, E, is set as follows: • E = 4 for CTU
• E = 2 for CE • E = 1 for CED • E = 3 for SD."
|
2022 |
Aggregate Statistics
ε: 10.0 |
Central | - | - | |
|
|
Movement Ranges Map
by Meta
|
Resulting in a total differential privacy epsilon value of 2.0.
Resulting in a total differential privacy epsilon value of 2.0.
|
2020 |
Aggregate-level
ε: 2.0 |
Central | - | - | |
|
|
On-device browser recommendations
by Brave
|
Brave is leveraging federated learning and differential privacy to provide on-device browser recommendations.
Brave is leveraging federated learning and differential privacy to provide on-device browser recommendations.
|
2021 |
User-level
|
Central | - | - | |
|
|
Private Third Party Audits
by Twitter, OpenMined
|
In September, 2022, in conjunction with the UN General Assembly and Christchurch Call leaders summit, New Zealand Prime Minister Jacinda Ardern and French President Emmanuel Macron announced the Christchurch Initiative on Algorithmic Outcomes, a partnership between New Zealand, the United States, Twitter, Microsoft and OpenMined to develop and test a differential privacy system to enable privacy preserving research across multiple online platforms.
In September, 2022, in conjunction with the UN General Assembly and Christchurch Call leaders summit, New Zealand Prime Minister Jacinda Ardern and French President Emmanuel Macron announced the Christchurch Initiative on Algorithmic Outcomes, a partnership between New Zealand, the United States, Twitter, Microsoft and OpenMined to develop and test a differential privacy system to enable privacy preserving research across multiple online platforms.
|
2022 |
Aggregate-level
|
Central | - | - | |
|
|
Energy Differential Privacy for OhmConnect Virtual Power Plant Measurement
by Recurve
|
Aggregate summary statistics describing how much energy OhmConnect participants saved during a demand response event (during the August 14, 2020 California Blackout), relative to a matched, non-participant group. Specifically, the statistics are:
- Average Load Shape: A private, hour-by-hour profile of the comparison group's average 'observed' and 'predicted' energy consumption over a 24-hour period.
- Percent Load Change: The percentage difference between total predicted and total observed energy consumption for each group during the 3-hour event window.
- Difference of Load Changes: The final net impact of the OhmConnect program, calculated by subtracting the comparison group's percent load change from the treatment group's percent load change.
Aggregate summary statistics describing how much energy OhmConnect participants saved during a demand response event (during the August 14, 2020 California Blackout), relative to a matched, non-participant group. Specifically, the statistics are:
- Average Load Shape: A private, hour-by-hour profile of the comparison group's average 'observed' and 'predicted' energy consumption over a 24-hour period.
- Percent Load Change: The percentage difference between total predicted and total observed energy consumption for each group during the 3-hour event window.
- Difference of Load Changes: The final net impact of the OhmConnect program, calculated by subtracting the comparison group's percent load change from the treatment group's percent load change.
|
2020 | Approximate DP |
user-level
ε: 4.72δ: 5.06e-09 |
Central | - | - |
|
|
Safari Energy Draining and Crashing Domains
by Apple
|
"For this use case, we set the parameters for HCMS to be m = 32,768, k = 1024, and \\(\epsilon\\) = 4 with p = 250,000 web domains."
"For this use case, we set the parameters for HCMS to be m = 32,768, k = 1024, and \\(\epsilon\\) = 4 with p = 250,000 web domains."
|
2017 | Pure DP |
User-level
ε: 4.0 |
Local | - | - |
|
|
Shared Mobility Dataset
by Google
|
The automated Laplace mechanism adds random noise drawn from a zero mean Laplace distribution and yields (ϵ, δ)-differential privacy guarantee of ϵ = 0.66 and δ = 2.1 × 10−29, which is very strong.
The automated Laplace mechanism adds random noise drawn from a zero mean Laplace distribution and yields (ϵ, δ)-differential privacy guarantee of ϵ = 0.66 and δ = 2.1 × 10−29, which is very strong.
|
2022 |
Aggregate Statistics
ε: 0.66δ: 2.1e-29 |
Central | - | - | |
|
|
Spanish Next Word Prediction
by Google
|
A machine learning model trained with federated learning for next‑word prediction on Spanish‑language Gboard.
A machine learning model trained with federated learning for next‑word prediction on Spanish‑language Gboard.
|
2022 | Zero-concentrated DP |
User-device-level
ρ: 0.81 |
Central | - | - |
|
|
Synthetic Data for Public Use
by Office of National Statistics UK
|
We recommend the use of generative adversarial networks (GANs) possibly in conjunction with privacy-preserving mechanisms such as differential privacy.
We recommend the use of generative adversarial networks (GANs) possibly in conjunction with privacy-preserving mechanisms such as differential privacy.
|
2020 |
Aggregate Statistics
|
Central | - | - | |
|
|
Telemetry Collection in Windows
by Microsoft
|
Data collection is performed every 6 hours, with \\(\epsilon\\) = 1.
Data collection is performed every 6 hours, with \\(\epsilon\\) = 1.
|
2017 |
User-level
ε: 1.0 |
Local | - | - | |
|
|
User-URL Privacy Dataset
by Meta
|
εuser = 1.453 for δ = 10−5.
εuser = 1.453 for δ = 10−5.
|
2020 |
User-level
ε: 1.453δ: 1.0e-05 |
Central | - | - | |
|
|
Vaccine Search Insights
by Google
|
The applied anonymization techniques protect every user’s daily search activity related to COVID-19 vaccinations with (ε, δ)-differential privacy for ε = 2.19 and δ = 10−5.
The applied anonymization techniques protect every user’s daily search activity related to COVID-19 vaccinations with (ε, δ)-differential privacy for ε = 2.19 and δ = 10−5.
|
2021 |
User-level
ε: 2.19δ: 1.0e-05 |
Central | - | - | |
|
|
Current Pageviews
by Wikimedia Foundation
|
Daily statistics of Wikipedia pageview counts, broken down by country of origin starting February 2023.
Daily statistics of Wikipedia pageview counts, broken down by country of origin starting February 2023.
|
2023 | Zero-concentrated DP |
Device-day
ρ: 0.015 |
Central | - | - |
|
|
Historical Pageviews
by Wikimedia Foundation
|
Daily statistics of Wikipedia pageview counts, broken down by country of origin between February 2017 and February 2023.
Daily statistics of Wikipedia pageview counts, broken down by country of origin between February 2017 and February 2023.
|
2023 | Pure DP |
Bounded contribution-day
ε: 1.0 |
Central | - | - |