For Immuta to enforce policies, it needs to catalog the resources policies are being applied to by performing metadata ingestion. Metadata ingestion is the process that occurs when you create an Immuta data source where Immuta gathers details about your tables. However, Immuta does not need access to the data within the tables in order to protect it, with the exception of a few specific and advanced masking policies detailed below.
Immuta collects and stores the following kinds of information in Immuta's Metadata Database for policy enforcement. Further, policy information may be transmitted to data source host systems for enforcement purposes as part of a query or to enable the host system to perform native enforcement.
Identity Management Information: Usernames, group information, and other kinds of personal identifiers may be stored and referenced for the purposes of performing authentication and access control and may be retained in audit logs. When such information is relevant for access determination under policy, it may be retained as part of the policy definition.
Schema Information: Data source metadata such as schema, column data types, and information about the host.
Immuta's Metadata Database can also contain the following forms of metadata for policy enforcement. These forms contain sample data from your tables and can be disabled if you do not want Immuta to have access to the data being protected.
Fingerprints: When enabled, additional statistical queries made during the health check are distilled into summary statistics, called fingerprints. During this process, statistical query results and data samples (which may contain PII) are temporarily held in memory by the Fingerprint Service.
k-Anonymization Policies: When a k-anonymization policy is applied, the columns under the k-anonymization policy are queried within a separate fingerprinting process which generates rules enforcing k-anonymity. The results of this query, which may contain PII, are temporarily held in memory by the Fingerprint Service. The final rules are stored for enforcement. Immuta requires that you opt in to use this masking policy type. To enable k-anonymization for your account, see the k-anonymization section on the app settings how-to guide.
Randomized Response Policies: If the list of substitution values for a categorical column is not part of the policy specification (e.g., when specified via the API), a list is obtained via query and merged into the policy definition.
If no metadata collection types have been disabled, data is processed in the following workflow to support data source creation, health checks, policy enforcement, and dictionary features.
A System Administrator configures the integration in Immuta.
A Data Owner registers data sources from their remote data platform with Immuta. Note: Data Owners can see sample data when editing a data source. However, this action requires the database password, and the small sample of data visible is only displayed in the UI and is not stored in Immuta.
When a data source is created or updated, the Metadata Database pulls in and stores statistics about the data source, including row count and high cardinality calculations.
The data source health check runs daily to ensure existing tables are still valid.
If an external catalog is enabled, the daily health check will pull in data source attributes (e.g., tags and definitions) and store them in the Metadata Database.
Immuta requires certain privileges to perform metadata ingestion. The user connecting a table to Immuta as a data source must have privileges specific to their data platform to perform metadata ingestion.
For example, a user registering a Snowflake table as an Immuta data source must have the REFERENCES privilege to view the structure of the table and allow Immuta access to that information as well. This does not require the user (or Immuta) to have access to view the data itself.
Policy decision data is transmitted to ensure end users querying data are limited to the appropriate access as defined by the policies in Immuta.
Spark plugin
In the Databricks Spark integration, the user, data source information, and query are sent to Immuta through the Spark plugin to determine what policies need to be applied while the query is being processed. Data that travels from Immuta to the Databricks cluster could include
user attributes.
what columns to mask.
the entire predicate itself (for row-level policies).
A user runs a query against data in their environment.
The query is sent to the Immuta Web Service.
The Web Service queries the Metadata Database to obtain the policy definition, which includes data source metadata (tags, column names, etc.) and user entitlements (groups and attributes).
The policy information is transmitted to the remote data system for native policy enforcement.
Query results are displayed based on what policy definition was applied.
Sample data is processed and aggregated or reduced during Immuta's fingerprinting process and specific policy processes. Note: Data Owners can see sample data when editing a data source. However, this action requires the database password, and the small sample of data visible is only displayed in the UI and is not stored in Immuta.
When enabled, statistical queries made during data source registration are distilled into summary statistics, called fingerprints. Fingerprinting allows Immuta to implement advanced privacy enhancing masking and data policies.
During this process, statistical query results and data samples (which may contain PII) are temporarily held in memory by the Fingerprint Service only for the amount of time it takes to calculate the statistics needed. For Snowflake, no data sample is needed, and only statistics about the data are returned to Immuta (no PII).
The fingerprinting process checks for new tables through schema monitoring (when enabled) and captures summary statistics of changes to data sources, including when policies were applied, external views were created, or sensitive data elements were added.
Immuta does not sample data for row redaction policies.
Immuta does not sample data for row redaction policies; Immuta only pulls samples of data to determine if a column is a candidate for randomized response and aggregates of user-defined cohorts for k-anonymization. Both datasets only exist in memory during the computation.
Sample data is processed when k-anonymization or randomized response policies are applied to data sources.
Sample data exists temporarily in memory in the Fingerprint Service during the computation.
k-Anonymization Policies: At the time of its application, the columns of a k-anonymization policy are queried under a separate fingerprinting process that generates rules enforcing k-anonymity. The results of this query, which may contain PII, are temporarily held in memory by the Fingerprint Service. The final rules are stored in the Metadata Database as the policy definition for enforcement. Immuta requires that you opt in to use this masking policy type. To enable k-anonymization for your account, see the k-anonymization section on the app settings how-to guide.
Randomized Response Policies: If the list of substitution values for a categorical column is not part of the policy specification (e.g., when specified via the API), a list is obtained via query and merged into the policy definition in the Metadata Database.
Raw data is processed for masking, producing either a distinct set of values or aggregated groups of values.
Immuta captures metadata and stores it in an internal PostgreSQL database. Customers can encrypt the volumes backing the database using an external Key Management Service to ensure that data is encrypted at rest.
To encrypt data in transit, Immuta uses TLS protocol, which is configured by the customer.
Immuta encrypts values with data encryption keys, either those that are system-generated or managed using an external key management service (KMS). Immuta recommends a KMS to encrypt or decrypt data keys and supports the AWS Key Management Service; however, if no KMS is configured, Immuta will generate a data encryption key on a user-defined rollover schedule, using the most recent data key to encrypt new values while preserving old data keys to decrypt old values.
Immuta employs three families of functions in its masking policies:
One-way Hashing: One-way (irreversible) hashing is performed via a salted SHA256 hash. A consistent salt is used for values throughout the data source, so users can count or track the specific values without revealing the true value. Since hashed values are different across data sources, users are unable to join on hashed values. Note: joining on masked values can be enabled in Immuta Projects.
Reversible Masking: For reversible masking, values are encrypted using AES-256 CBC encryption. Encryption is performed using a cell-specific initialization vector. The resulting values can be unmasked by an authorized user. Note that this is dynamic encryption of individual fields as results are streamed to the querying system; Immuta is not modifying records in the data store.
Reversible Format Preserving Masking: Format preserving masking maintains the format of the data while masking the value, and is achieved by initializing and applying the NIST standard method FF1 at the column level. The resulting values can be unmasked by an authorized user.
Immuta communicates with remote databases over a TCP connection.