Data Minimization

Before any data collection begins, the specific, explicit, and legitimate purpose must be documented. Data cannot be repurposed for incompatible secondary uses without new consent or legal basis.

• Purpose Binding: Data collected for analytics cannot be silently diverted to train a machine learning model.
• Granular Consent: Users must be able to consent to specific processing activities independently.
• Article 5(1)(b) GDPR: The legal codification of this principle in European law.

The three-part test for every data element collected. Adequacy ensures sufficient data to fulfill the purpose. Relevance requires a rational link between the data and the purpose. Necessity demands that the purpose cannot be achieved with less data.

• Example: A newsletter signup requires an email address (necessary) but not a phone number or date of birth (unnecessary).
• Golden Rule: If you cannot articulate why a field is strictly required, do not collect it.

Personal data must be kept in a form that permits identification of data subjects for no longer than necessary for the purposes for which it is processed. This mandates automated data retention schedules and secure deletion mechanisms.

• Anonymization: Data may be stored indefinitely only if truly anonymized and re-identification is impossible.
• Automated TTLs: Implement time-to-live policies in databases to auto-purge records after a defined retention period.
• Pseudonymization: Separating direct identifiers from operational data to reduce risk during active processing.

Systems must be architected so that, by default, only the minimum necessary data is processed. This shifts the burden from user action to system design.

• Privacy-Preserving Defaults: The strictest privacy settings must be the factory default, not an opt-in.
• Client-Side Processing: Perform computations locally on the user's device so raw data never leaves their control.
• Just-in-Time Collection: Request data only at the moment it is needed, not preemptively during onboarding.

Instead of collecting individual-level data, derive insights from aggregated statistics or pre-computed summaries. This technique satisfies analytical requirements without exposing personal records.

• Differential Privacy: Inject calibrated noise into aggregate queries to mathematically guarantee individual privacy.
• On-Device Federated Analytics: Compute aggregate metrics locally and upload only the anonymized summary.
• K-Anonymity: Ensure any released aggregate is indistinguishable from at least k-1 other individuals.

Replace real personal data with high-fidelity synthetic data generated by models like GANs or VAEs. The artificial data preserves statistical properties for development and testing without exposing actual records.

• Private Synthetic Data Factories: Generate synthetic datasets on-premises within a Trusted Execution Environment.
• Zero Real Exposure: Developers and QA teams interact only with artificial data, never production records.
• Referential Integrity: Multi-table synthesis preserves foreign key relationships across generated datasets.

A mathematical framework that injects calibrated statistical noise into datasets or queries to guarantee that the presence or absence of any single individual's record is indistinguishable.

• Quantified by the privacy budget (epsilon) — lower values enforce stronger guarantees
• Prevents re-identification even when attackers possess auxiliary information
• Implemented via mechanisms like the Laplace or Gaussian noise addition

A privacy property ensuring that each released record is indistinguishable from at least k-1 other records with respect to quasi-identifier attributes (e.g., ZIP code, age, gender).

• Prevents singling out of individuals in a dataset
• Achieved through generalization (replacing specific values with ranges) and suppression (removing outlier records)
• Limitations: vulnerable to homogeneity attacks and background knowledge attacks if sensitive attributes lack diversity

The processing of personal data to replace direct identifiers (name, email, social security number) with artificial pseudonyms or tokens.

• Unlike anonymization, the data remains indirectly identifiable if the mapping table is preserved
• Requires separate storage of the re-identification key with strict access controls
• Explicitly recognized under GDPR Article 4(5) as a risk-reduction measure, not a full anonymization technique

The process of obscuring specific sensitive data elements by replacing them with structurally similar but inauthentic characters or values.

• Static masking: creates a permanently sanitized copy for non-production environments
• Dynamic masking: obscures data in real-time based on user privileges without altering the underlying store
• Common techniques include character shuffling, substitution, and nullification
• Often used as a preprocessing step before feeding data into synthetic generators

A privacy attack where an adversary determines whether a specific data record was included in a model's training set by analyzing its output behavior.

• Exploits the tendency of models to be overconfident on memorized training examples
• Shadow model technique: attacker trains local models to mimic target behavior and learns to distinguish members from non-members
• Mitigation strategies include DP-SGD, regularization, and limiting overfitting
• Directly validates whether data minimization and privacy-preserving training are effective

A quantifiable limit on the total privacy loss allowed over a series of queries or data releases, parameterized by epsilon (ε).

• Lower epsilon = stronger privacy, more noise injected
• Composition theorems track cumulative loss across multiple analyses
• Typical values: ε = 0.1 to 1 for high-privacy scenarios, ε = 1 to 10 for balanced utility
• Once the budget is exhausted, no further queries on that dataset are permitted
• Central to enforcing data minimization in interactive query systems