Prepare and test your skills
Prepare and test your skills
Differentiate Between Anonymization and Pseudonymization means comparing two ways to hide personal data while keeping it useful for AI development. Pseudonymization replaces a direct identifier with a reversible stand-in so records can still be linked. Anonymization removes or masks sensitive values so they cannot be linked back to a person.
Data risk management is the practice of storing, processing, and using data within a risk level that fits the business. In Sensitive Data Protection (opens in a new tab), which includes Cloud Data Loss Prevention (Cloud DLP), the goals are to discover and classify data, understand exposure risk, and protect data with controls or de-risk it through obfuscation:
When a workload is evaluated, three questions define the privacy risk: what kind of data the workload handles and whether any of it is sensitive, whether the data is exposed to the right users in the right environment for an approved purpose, and whether risk can be reduced through data minimization and obfuscation. Sharing data with more people and more development environments increases exposure, because the data is stored and processed in more places.
Pseudonymization is the technique that hides the original identifier but preserves referential integrity so analysis can still group records by the same person. Data flows from the collecting team to the analytics team to the machine learning model, and pseudonymization sits in between to reduce risk across that trust boundary. A common example is customer feedback for a model: the structured personally identifiable information (PII), such as the customer email address used as a primary key, is replaced through tokenization. The analytics team never sees the actual email, yet the token stays consistent so the team can still tell which reviews belong to the same customer and aggregate per customer.
Because the same token always stands for the same person, utility for training and analysis stays high. Re-identification risk is lowered but not removed, since someone with access to the original mapping could reverse the tokens. Pseudonymization therefore depends on access controls that keep the mapping separate from the tokenized dataset and limit who can see each side.
Anonymization is the technique that inspects data and masks or removes sensitive values so the original person cannot be recovered from the shared dataset. Freeform product reviews are a clear example, because unstructured data can contain intermittent instances of sensitive data like phone numbers and addresses. The protection service scans the review text, finds those values, and masks them before the dataset is shared. The analytics team receives the meaning of the feedback without receiving the phone number or address inside it.
This approach lowers re-identification risk more strongly than pseudonymization because there is no token to reverse for those masked values. The tradeoff is lower utility for identity-based grouping, since a masked phone number cannot be used to join records. Choose anonymization when the analytics team needs the content of the data but does not need to know who wrote it.
Comparing the two approaches for AI data shows how each method balances re-identification risk against data utility for model development:
A comparison matrix contrasting pseudonymization and anonymization across four dimensions: technique applied, what happens to the data, utility preserved, and re-identification risk, ending with when to choose each approach.
| Approach | What happens to the data | Utility preserved | Re-identification risk |
|---|---|---|---|
| Pseudonymization with tokenization | Structured PII such as customer email is replaced with a consistent token | High for per-customer aggregation, because reviews from the same customer still link together | Reduced, but reversible if the token mapping is exposed |
| Anonymization with masking | Intermittent sensitive data in unstructured text such as phone numbers and addresses is masked | High for content analysis, but low for linking records by identity | Lower, because the masked value cannot be recovered from the shared data |
Exam tip: Tokenize structured identifiers when you must keep referential integrity across records; mask unstructured findings when you must remove them entirely.
Identify and Mitigate Privacy Risks in Generative AI means recognizing how sensitive information can leak through prompts, outputs, and training data, and knowing the controls that limit that exposure. Generative AI security (opens in a new tab) in Google Cloud limits who can access customer data, logs administrative access, encrypts data at rest and in use, and detects and responds to access violations.
Privacy risks in generative AI workflows appear at several points where sensitive data moves through the system. Sensitive information can be exposed when it is included in model prompts, when it appears in model outputs, or when it is memorized from training data during fine-tuning. Google Cloud's data usage commitments state that customer data is not used to train AI models without permission, is not scanned for advertising, and is not sold to third parties. Administrative access to customer data is restricted to a small group of employees based on job function, using least privilege and need-to-know principles.
Services that monitor and control Google's access to data give customers visibility into administrative access and control over encryption keys and regulated workloads:
| Service | What it does |
|---|---|
| Access Approval (opens in a new tab) | Requires your approval before an authorized Google administrator can access your data to support you. Approved requests are logged with Access Transparency logs. |
| Access Transparency | Logs administrative access by Google authorized personnel when they support your organization or maintain service availability. |
| Assured Workloads | Creates and monitors control packages for regulatory programs such as FedRAMP, ITAR, or Sovereign Controls for EU. |
| Cloud KMS with Cloud EKM | Encrypts data with keys stored and managed in a third-party key management system deployed outside Google's infrastructure. |
| Confidential Computing | Encrypts data in use. |
Data access and restrictions for customer data start from the rule that customer data stored on Google systems belongs to the customer. Google does not scan data for advertising, sell it to third parties, or use it to train AI models without permission. Administrative access is logically isolated per customer even when data is stored on shared hardware, and requests for additional access follow a formal approval process managed by workflow tools that keep audit records. Support services are provided only to authorized customer administrators, and audit logs are available through Access Transparency.
Incident detection and response process is the ordered workflow Google uses to find and fix suspected access violations with threat intelligence and automated controls:
Compliance support and verification means Google Cloud undergoes independent checks of its security, privacy, and compliance controls. Key international standards include ISO/IEC 27001 for Information Security Management, ISO/IEC 27017 for Cloud Security, ISO/IEC 27018 for Cloud Privacy, and ISO/IEC 27701 for Privacy. SOC 2 and SOC 3 reports are also available, and sector-specific frameworks include FedRAMP, BSI C5, and MTCS.
Apply Google Cloud Privacy Tools and Controls means using services and settings to find sensitive data, hide it, and limit who can reach it. In a generative AI workflow (opens in a new tab), these controls apply at every stage: scanning training or prompt data for personally identifiable information (PII), masking that data before it reaches a model or a user, and logging or approving any administrative access to the underlying data.
Discovering and masking sensitive data starts with Sensitive Data Protection, also known as Cloud Data Loss Prevention (Cloud DLP). Sensitive Data Protection scans data for sensitive and high-risk information such as PII, financial data, and health information. It is the discovery step: before data is used in a generative AI pipeline, the scan identifies which fields or records contain regulated content. Once sensitive data is found, data masking obscures it in place while still letting authorized users work with the surrounding, non-sensitive data. Masking can also target values that match sensitive data patterns, which protects against accidental disclosure when data is shared or displayed.
BigQuery adds several related controls that work alongside Sensitive Data Protection:
Controlling Google personnel access to customer data gives customers visibility into, and control over, when Google personnel can reach their data. These controls matter for generative AI because training data, prompts, and outputs often contain the most sensitive content an organization holds.
Exam tip: Access Transparency records administrative access after it happens, while Access Approval blocks that access until the customer approves the request first.
Controlling access within your organization applies least privilege and need-to-know so access matches defined responsibilities. Predefined roles can be assigned to users, and custom roles can be created when more granular permissions are needed. Row-level security and column-level security restrict access to specific rows and columns in a table, and VPC Service Controls establishes an access perimeter around Google Cloud resources so data cannot move outside a defined boundary.
For sharing, BigQuery uses live, zero-copy data sharing: data is shared in place with no integration, data movement, or replication, and linked datasets act as a live pointer to the shared asset. Data owners keep control over their data and can define rules or policies that restrict access and usage. When organizations need to collaborate on sensitive data, data clean rooms provide a controlled environment for that work.
Google data usage commitments state that data stored on its systems belongs to the customer. It is not scanned for advertising purposes, not sold to third parties, and not used to train AI models without permission. The Data Processing Addendum commits Google to processing data only to meet contractual obligations, and customers who stop using the service can take their data with them without penalty or additional cost. Internally, Google logically isolates each customer's data from other customers even when it sits on the same physical server, and only a small group of employees have access to customer data, granted through a formal request-and-approval process that is tracked by workflow tools with audit records.
Pseudonymization replaces a direct identifier such as a customer email address with a reversible stand-in token that stays consistent so records can still be linked by the same person. Anonymization removes or masks sensitive values such as phone numbers and addresses in unstructured text so they cannot be linked back to a person from the shared dataset.
Use pseudonymization with tokenization for structured personally identifiable information (PII) such as customer email when you must keep referential integrity for per-customer aggregation. Use anonymization with masking for intermittent sensitive data in unstructured text such as phone numbers and addresses when the analytics team needs the content of the data but does not need to know who wrote it.
Access Transparency logs administrative access by authorized Google personnel after it happens when they support an organization or maintain service availability. Access Approval blocks that access until the customer gives explicit approval first, and approved requests are logged with Access Transparency logs.
Sensitive Data Protection, which includes Cloud Data Loss Prevention (Cloud DLP), scans data for sensitive and high-risk information such as personally identifiable information (PII), financial data, and health information to discover and classify it. Once sensitive data is found, data masking obscures it in place while still letting authorized users work with the surrounding non-sensitive data.
An enterprise customer service department is deploying a generative AI virtual assistant to handle complex customer inquiries. To improve response accuracy, the engineering team fine-tunes a foundational large language model using historical support tickets, email logs, and customer chat transcripts containing resolution details.
During security testing, engineers discover that specific prompt queries cause the model to output verbatim excerpts of unredacted customer names, home addresses, and private account details that were present in the historical support transcripts.
Which generative AI privacy vulnerability does this scenario demonstrate?
Prompt token truncation
Model hallucination
Data drift and distribution shift
Training data memorization and regurgitation
Last updated