Data Minimization

Data must be collected for specified, explicit, and legitimate purposes and not further processed in a manner incompatible with those purposes. This is the legal and architectural anchor of minimization.

• Proactive declaration: Define the exact ML objective before ingestion begins.
• Incompatibility detection: Automated pipelines flag when data is repurposed for a new, undeclared training task.
• Example: A fraud detection model ingests transaction amount and merchant category but excludes the purchaser's full name and home address because they are irrelevant to the anomaly detection purpose.

Only data that is adequate (sufficient to fulfill the purpose), relevant (has a rational link to the purpose), and necessary (the purpose cannot be fulfilled without it) may be processed.

• Feature ablation testing: Remove a feature and measure model performance; if accuracy doesn't drop, the feature was unnecessary.
• Granularity reduction: Replace precise timestamps with hour-of-day buckets or exact GPS coordinates with municipality-level geohashes.
• Example: Training a churn predictor requires recency and frequency of logins, but not the specific IP addresses or device fingerprint hashes.

Data must be kept in a form that permits identification for no longer than necessary. This requires automated, verifiable data purging mechanisms.

• Time-to-live (TTL) policies: Training datasets are automatically tombstoned and purged after a defined period.
• Ephemeral compute: Raw sensitive data is processed in-memory and never persisted to disk, aligning with transient processing principles.
• Example: A speech recognition model retains raw audio for 72 hours for quality assurance, then automatically deletes it, keeping only the anonymized text transcript.

Data should be aggregated to a level where individual records are statistically indistinguishable. K-anonymity ensures each released record is identical to at least k-1 other records.

• Suppression: Redact rare attribute combinations that create unique identifiers in the dataset.
• Generalization: Replace specific values with broader categories (e.g., age 37 → age range 30-40).
• Example: A healthcare analytics dashboard only displays condition prevalence statistics when the cohort size exceeds k=50 patients, preventing singling out individuals with rare diseases.

Process raw data directly on the user's device and transmit only the minimized, abstracted model updates or inferences to central servers.

• On-device inference: The model runs locally; only the final classification label leaves the device.
• Federated learning integration: Raw data never leaves the edge node; only encrypted, clipped gradient updates are shared.
• Example: A keyboard's next-word prediction model trains on-device. The raw typed text is never uploaded; only a differentially private gradient vector is sent to the aggregation server.

Architect the ingestion pipeline to actively discard non-essential fields at the earliest possible stage, before data enters the training environment.

• Schema-on-write enforcement: Ingestion schemas strictly whitelist permitted fields; all extraneous JSON keys are dropped silently.
• Tokenization at ingestion: Replace direct identifiers (email, phone) with salted, irreversible pseudonyms immediately upon receipt.
• Example: An API gateway for a recommendation engine strips HTTP headers, browser user-agent strings, and referrer URLs, forwarding only the product ID and session token to the ML serving layer.