Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 4 additions & 11 deletions deployments/apple_iconic_scenes_photo_selections_2023.yaml
Original file line number Diff line number Diff line change
@@ -1,6 +1,4 @@
url_slug: iconic-scenes-and-photo-sections-for-ios-apps-apple-2023
status: Draft
tier: 3
deployment:
basic:
# Tier 1:
Expand All @@ -9,12 +7,12 @@ deployment:
-
data_curator_name: 'Apple'
data_curator_type: 'Industry'
description: This deployment identifies frequently photographed iconic locations across users, enabling Apple’s Photos app to automatically select representative images for features such as Memories. The system is designed to protect individual user data while providing personalized content.
intended_use: Identify frequently photographed iconic locations across users, enabling Apple’s Photos app to automatically select representative images for features such as Memories, all while protecting individual user data.
data_product_type: Machine learning model
data_product_region: Global
description: This deployment identifies frequently photographed iconic locations across users, enabling Apple’s Photos app to automatically select representative images for features such as Memories. The system is designed to protect individual user data while providing personalized content.
publication_date: '2023-07-21'
data_product_sector: Technology
publication_date: '2023-07-21'

# Tiers 2 and 3:
dp_variant:
Expand All @@ -32,19 +30,12 @@ deployment:

deployment_model:
model_name: Local
model_name_details: Local Differential Privacy
# actors: ''
release_type: Many releases
release_type_details: Image data is collected from users who opted in to the Analytics & Improvements feature on their Apple devices and be used for improving features in Apple’s Photos app.
data_source_type: Dynamic
# data_source_type_details: ''
access_type: Non-interactive
# access_type_details: ''

# accounting:
# post_processing: '' # Tier 3
# composition: '' # Tier 3

implementation:
# preprocessing_and_hyperparameter_tuning: '' # Tier 3
mechanisms: |
Expand All @@ -58,5 +49,7 @@ deployment:
sources: |
- Blog post: https://machinelearning.apple.com/research/scenes-differential-privacy
# notes: '' # other extra information
status: Draft
registry_authors:
- Michael Khavkin
tier: 3
30 changes: 13 additions & 17 deletions deployments/apple_popular_emojis.yaml
Original file line number Diff line number Diff line change
@@ -1,19 +1,17 @@
url_slug: popular-emojis-apple-2017
status: Pending
tier: 3
deployment:
basic:
name: Popular Emojis
data_curators:
-
data_curator_name: 'Apple'
data_curator_type: 'Industry'
description: Summary statistics of frequency histograms of emoji usage per locale.
intended_use: For measurement (within Apple) of which emojis are most used, and their relative frequencies, across keyboard locales, to improve QuickType’s predictive emoji suggestions.
data_product_type: Summary statistics
data_product_region: Global
description: Summary statistics of frequency histograms of emoji usage per locale.
publication_date: '2017-01-01'
data_product_sector: Technology
publication_date: '2017-01-01'

dp_variant:
variant_name: Pure DP
Expand Down Expand Up @@ -47,15 +45,6 @@ deployment:
access_type: Non-interactive
# access_type_details: ''

accounting:
post_processing: |
- Server discards IP addresses associated with each record.
- Aggregation and Sketch Matrix Construction: The server aggregates the anonymized, private records into a sketch matrix, which is a summary grid with 65,536 rows and 1,024 columns. Each row corresponds to a possible hash function that the user’s device could have used. For each private report received, the server uses the included index to identify which row to update , and then adds the user's noisy data to that specific row. The random noise tends to average out, allowing the underlying patterns of popular emojis to be estimated in the next step.
- Frequency Estimation and Debiasing: The server estimates the frequency of each emoji in the dictionary based on the sketch matrix, averages these values, and debiases the result, producing a final estimated count for that emoji.
- Final Histogram Generation and Thresholding: The estimated counts for all emojis are assembled into a histogram, and thresholding is performed so that only emojis whose estimated counts are above a predetermined threshold are included in the final data product.

composition: No information was provided about composition.

implementation:
preprocessing_and_hyperparameter_tuning: 'Three parameters are explicitly set: \\(\epsilon = 4\\), number of hash functions k = 65,546 and the size of the privatized vector m = 1024. These were predetermined by the authors based on an analysis of the tradeoff between utility, server computation cost (related to k), and device bandwidth (related to m).'
mechanisms: |
Expand All @@ -69,18 +58,25 @@ deployment:
- The privatized records are temporarily stored on-device using data protection, rather than being immediately transmitted to the server. After a delay based on device conditions, the system randomly samples from the differentially private records subject to a predetermined limit on the number of privatized records that can be transmitted daily for each use case, and sends the sampled records to the server. These records do not include device identifiers or timestamps of when events were generated.
- After the data is privatized, the resulting records are not sent instantaneously, but are stored on-device using data protection and sent as part of a random batch once a day.
- The privatized records are sent to the server with TLS encryption.

composition: No information was provided about composition.
post_processing: |
- Server discards IP addresses associated with each record.
- Aggregation and Sketch Matrix Construction: The server aggregates the anonymized, private records into a sketch matrix, which is a summary grid with 65,536 rows and 1,024 columns. Each row corresponds to a possible hash function that the user’s device could have used. For each private report received, the server uses the included index to identify which row to update , and then adds the user's noisy data to that specific row. The random noise tends to average out, allowing the underlying patterns of popular emojis to be estimated in the next step.
- Frequency Estimation and Debiasing: The server estimates the frequency of each emoji in the dictionary based on the sketch matrix, averages these values, and debiases the result, producing a final estimated count for that emoji.
- Final Histogram Generation and Thresholding: The estimated counts for all emojis are assembled into a histogram, and thresholding is performed so that only emojis whose estimated counts are above a predetermined threshold are included in the final data product.
justification: |
- The choice of \\(\epsilon\\) was “based on the privacy characteristics of the underlying dataset” and is “consistent with the parameters proposed in the differential privacy research community”.
- The CMS algorithm yields a high number of hash collisions (mapping 2,600 emojis to 1024 bits) providing further plausible deniability.
additional_information: |
- Under "Security Measures": "The privatized records are temporarily stored on-device using data protection"; the paper points to this link https://www.apple.com/business/docs/iOS_Security_Guide.pdf
- The paper covers collecting aggregate statistics for different use cases: trending new words, popular emojis, popular HealthKit data types, Safari media-playback preferences, and high energy/memory domains. We only treated the popular emojis use case, treating it as one deployment.

administrative:
sources: |
- Paper: https://docs-assets.developer.apple.com/ml-research/papers/learning-with-privacy-at-scale.pdf
notes: |
- Under "Security Measures": "The privatized records are temporarily stored on-device using data protection"; the paper points to this link https://www.apple.com/business/docs/iOS_Security_Guide.pdf
- The paper covers collecting aggregate statistics for different use cases: trending new words, popular emojis, popular HealthKit data types, Safari media-playback preferences, and high energy/memory domains. We only treated the popular emojis use case, treating it as one deployment.
status: Pending
registry_authors:
- Elena Ghazi
- Nicolas Berrios
- Jack Fitzsimons
tier: 3
13 changes: 7 additions & 6 deletions deployments/assistive_ai.yaml
Original file line number Diff line number Diff line change
@@ -1,19 +1,17 @@
url_slug: assistive-ai-microsoft-2020
status: Pending
tier: 2
deployment:
basic:
name: Assistive AI
data_curators:
-
data_curator_name: 'Microsoft'
data_curator_type: 'Industry'
description: Privately refining the set of responses suggested for email and chat.
intended_use: The AI system was trained internally at Microsoft to be deployed in their products. The model provides suggestions for replies to emails and texts, and is intended to make responding to emails and texts easier for users. The underlying data are “hundreds of millions of Messages (Emails or chats) and their Replies called Message-Reply (MR) pairs.”
data_product_type: Machine learning model
data_product_region: Global
description: Privately refining the set of responses suggested for email and chat.
publication_date: '2020-01-01'
data_product_sector: Technology
publication_date: '2020-01-01'

dp_variant:
variant_name: Approximate DP
Expand All @@ -32,14 +30,17 @@ deployment:
preprocessing_and_hyperparameter_tuning: '' # TODO: Fill in correct value
mechanisms: Differentially Private Set Union
justification: '' # TODO: Fill in correct value
additional_information: |
- Restricted to some languages (EN, ES, FR, PR, IT)

administrative:
sources: |
- Blog post: https://www.microsoft.com/en-us/research/articles/assistive-ai-makes-replying-easier-2/
- Paper: https://arxiv.org/pdf/2002.09745
notes: |
- Restricted to some languages (EN, ES, FR, PR, IT)
status: Pending
registry_authors:
- Priyanka Nanayakkara
- Nicolas Berrios
- Jack Fitzsimons
- James Honaker
tier: 2
13 changes: 7 additions & 6 deletions deployments/audience_engagement.yaml
Original file line number Diff line number Diff line change
@@ -1,19 +1,17 @@
url_slug: audience-engagement-api-linkedin-2020
status: Pending
tier: 1
deployment:
basic:
name: Audience Engagement API
data_curators:
-
data_curator_name: 'LinkedIn'
data_curator_type: 'Industry'
description: Aggregated summary statistics about content engagement with LinkedIn posts.
intended_use: For data analysts to learn about the LinkedIn user base. For example, marketers may want to learn about the kind of content LinkedIn users engage with before they create sponsored content.
data_product_type: Interactive API
data_product_region: Global
description: Aggregated summary statistics about content engagement with LinkedIn posts.
publication_date: '2020-01-01'
data_product_sector: Technology
publication_date: '2020-01-01'
dp_variant:
variant_name: Approximate DP
data_domain: '' # TODO
Expand All @@ -36,13 +34,16 @@ deployment:
preprocessing_and_hyperparameter_tuning: '' # TODO: Fill in correct value
mechanisms: Laplace, Gumbel
justification: '' # TODO: Fill in correct value
additional_information: |
- The paper states that “data is purged within 30 days of a user leaving the ecosystem because it has a 30 day retention.” We take this to mean that the data are refreshed on a monthly basis, i.e., that leaving the ecosystem means leaving a single user session on LinkedIn as opposed to account deletion.

administrative:
sources: |
- Paper: https://arxiv.org/pdf/2002.05839
notes: |
- The paper states that “data is purged within 30 days of a user leaving the ecosystem because it has a 30 day retention.” We take this to mean that the data are refreshed on a monthly basis, i.e., that leaving the ecosystem means leaving a single user session on LinkedIn as opposed to account deletion.
status: Pending
registry_authors:
- Priyanka Nanayakkara
- Nicolas Berrios
- Jack Fitzsimons
- James Honaker
tier: 1
8 changes: 4 additions & 4 deletions deployments/autoplay_intent.yaml
Original file line number Diff line number Diff line change
@@ -1,19 +1,17 @@
url_slug: autoplay-intent-apple-2017
status: Pending
tier: 1
deployment:
basic:
name: AutoPlay Intent
data_curators:
-
data_curator_name: 'Apple'
data_curator_type: 'Industry'
description: Identify which websites users typically prefer to have auto-play with sound for video elements.
intended_use: Video Browser Play
data_product_type: Summary statistics # TODO: Is this correct?
data_product_region: Not Specified
description: Identify which websites users typically prefer to have auto-play with sound for video elements.
publication_date: '2017-01-01'
data_product_sector: Technology
publication_date: '2017-01-01'
dp_variant:
variant_name: Pure DP
data_domain: '' # TODO
Expand All @@ -36,7 +34,9 @@ deployment:
justification: '' # TODO: Fill in correct value
administrative:
sources: https://docs-assets.developer.apple.com/ml-research/papers/learning-with-privacy-at-scale.pdf
status: Pending
registry_authors:
- Nicolas Berrios
- Jack Fitzsimons
- James Honaker
tier: 1
9 changes: 4 additions & 5 deletions deployments/birth_dataset.yaml
Original file line number Diff line number Diff line change
@@ -1,19 +1,17 @@
url_slug: israel-s-national-registry-of-live-births-of-2014-israeli-ministry-of-health-2024
status: Pending
tier: 1
deployment:
basic:
name: Israel's National Registry of Live Births of 2014
data_curators:
-
data_curator_name: 'Israeli Ministry of Health'
data_curator_type: 'Government'
description: Synthetic data covering live births in Israel during 2014.
intended_use: Health Data
data_product_type: Synthetic data
data_product_region: Israel
description: Synthetic data covering live births in Israel during 2014.
publication_date: '2024-02-01'
data_product_sector: Healthcare
publication_date: '2024-02-01'
dp_variant:
variant_name: Pure DP
privacy_loss:
Expand All @@ -28,7 +26,6 @@ deployment:
is \\(\epsilon = 9.98\\)."
deployment_model:
model_name: Central
# model_name_details: TODO
release_type: One release
release_type_details: Static
access_type: Non-interactive
Expand All @@ -38,6 +35,8 @@ deployment:
justification: '' # TODO: Fill in correct value
administrative:
sources: https://birth.dataset.pub
status: Pending
registry_authors:
- Shlomi Hod
- James Honaker
tier: 1
35 changes: 18 additions & 17 deletions deployments/broadband_coverage.yaml
Original file line number Diff line number Diff line change
@@ -1,19 +1,17 @@
url_slug: broadband-coverage-microsoft-2021
status: Pending
tier: 3
deployment:
basic:
name: Broadband Coverage
data_curators:
-
data_curator_name: 'Microsoft'
data_curator_type: 'Industry'
description: Publicly available U.S. Broadband Coverage dataset that reports broadband coverage percentages at a zip code-level.
intended_use: Allow researchers and policy makers to understand problems related to broadband access and develop solutions to improve broadband access while simultaneously preserving the privacy of individual households.
data_product_type: Dataset
data_product_region: USA
description: Publicly available U.S. Broadband Coverage dataset that reports broadband coverage percentages at a zip code-level.
publication_date: '2021-04-01'
data_product_sector: Technology # Telecommunication, Public Sector/Government?
publication_date: '2021-04-01'

dp_variant:
variant_name: Pure DP
Expand All @@ -30,7 +28,6 @@ deployment:

deployment_model:
model_name: Central
# model_name_details: ''
actors: |
- Users whose devices connect to Microsoft services. They must trust Microsoft to handle their data and apply the DP protections as described.
- Microsoft (Curator): collects the raw telemetry data, applies the Laplace mechanism to produce differentially private counts, and publishes the final data product.
Expand All @@ -42,10 +39,15 @@ deployment:
access_type: Non-interactive
# access_type_details: ''

accounting:
post_processing: |
- Negative values are clamped to zero.
- The final privatized Broadband Coverage Estimation was calculated using noisy counts and the number of households in each ZIP code as reported in the Department of Housing and Urban Development’s publicly available dataset.
implementation:
preprocessing_and_hyperparameter_tuning: The 32,653 zip codes were predefined based on the average number of machines in the last years (not on the dataset being protected).

mechanisms: |
- The Laplace mechanism with \\(\epsilon = 0.1\\) was applied to every count query (see section on Composition for query descriptions).
- All DP processing was done with the OpenDP SmartNoise library: https://github.com/microsoft/USBroadbandUsagePercentages
- No information was provided about protection against floating point errors.
- No information was provided about leakage due to idiosyncrasies of the computing platform.
- No information was provided about security measures in place to protect the underlying data.

composition: |
The following quantities are computed with DP (per ZIP code):
Expand All @@ -61,15 +63,12 @@ deployment:
- The total privacy loss corresponding to Microsoft services data is 0.1.
By sequential composition, the total privacy loss is 0.2.
(Note: The equation for Broadband Coverage Estimate does not use the "number of devices with speed < 25". In fact, it seems this number is never used.)
implementation:
preprocessing_and_hyperparameter_tuning: The 32,653 zip codes were predefined based on the average number of machines in the last years (not on the dataset being protected).

mechanisms: |
- The Laplace mechanism with \\(\epsilon = 0.1\\) was applied to every count query (see section on Composition for query descriptions).
- All DP processing was done with the OpenDP SmartNoise library: https://github.com/microsoft/USBroadbandUsagePercentages
- No information was provided about protection against floating point errors.
- No information was provided about leakage due to idiosyncrasies of the computing platform.
- No information was provided about security measures in place to protect the underlying data.
post_processing: |
- Negative values are clamped to zero.
- The final privatized Broadband Coverage Estimation was calculated using noisy counts and the number of households in each ZIP code as reported in the Department of Housing and Urban Development’s publicly available dataset.


justification: |
- The privacy budget \\(\epsilon = 0.1\\) used to obtain the counts is chosen to be comparable to other industry differentially private data releases.
- The number of devices connected to the internet at broadband speed per each zip code is counted based on the Federal Communications Commission (FCC)’s definition of broadband that is 25mbps per download.
Expand All @@ -79,7 +78,9 @@ deployment:
sources: |
- Paper: https://arxiv.org/abs/2103.14035
- Github: https://github.com/microsoft/USBroadbandUsagePercentages
status: Pending
registry_authors:
- Elena Ghazi
- Nicolas Berrios
- Jack Fitzsimons
tier: 3
Loading
Loading