Privacy by Design (PbD)

This principle mandates anticipating and preventing privacy-invasive events before they occur. It rejects the 'fix it later' mentality, requiring engineers to design systems that actively avoid data breaches and privacy harms. In practice, this means conducting threat modeling for privacy during the architecture phase, not waiting for a post-breach audit. It shifts the posture from damage control to risk elimination.

The individual's privacy must be automatically protected without requiring any manual action. This is the 'no action required' principle.

• Purpose Specification: Data collection must be limited to the stated purpose.
• Collection Limitation: Collect only the minimum necessary data.
• Use Limitation: Data is not used for undisclosed secondary purposes.
• Retention Limitation: Data is deleted when no longer needed. This ensures that even if a user ignores settings, their privacy remains intact.

Privacy cannot be a bolt-on feature; it must be a core component of the system's architecture. This means integrating privacy controls directly into the code, data structures, and APIs from the ground up. For ML engineers, this translates to embedding techniques like on-device federated learning or differential privacy noise injection directly into the training pipeline, rather than masking data after the fact.

This principle rejects the false dichotomy between privacy and security/usability. It advocates for a positive-sum paradigm where all objectives are accommodated. Engineers must find creative solutions that deliver robust functionality and strong privacy simultaneously. For example, homomorphic encryption allows computation on encrypted data, delivering analytical insights without sacrificing confidentiality.

Data must be securely managed throughout its entire lifecycle, from collection to destruction. This requires cradle-to-grave protection.

• Confidentiality: Strong encryption at rest and in transit.
• Integrity: Ensuring data is not altered maliciously.
• Availability: Resilient systems.
• Secure Destruction: Cryptographic shredding and secure deletion protocols at the end of the retention period.

Systems must operate according to stated promises and be independently verifiable. Component parts and operations must remain visible to users and providers. This builds algorithmic trust. For ML systems, this involves publishing model cards, documenting training data provenance, and implementing explainability tools (like SHAP or LIME) that allow users to understand how decisions are made.

Architects must prioritize the interests of the individual by offering strong defaults, appropriate notice, and user-friendly options. This is the human-centric anchor of the framework. It requires empowering users with granular consent management and self-service data access tools. The design must be empathetic to the user's need for autonomy over their personal information.

The core PbD principle of limiting collection to strictly necessary attributes. Data minimization reduces the attack surface by ensuring that data not collected cannot be breached.

• Just-in-time collection: Gathering data only at the moment of need
• Aggregation: Using summary statistics instead of raw records
• Retention limits: Automating deletion when the processing purpose expires

A mathematical framework providing provable privacy guarantees by injecting calibrated noise into query results. It ensures the output distribution is nearly identical whether or not any single individual is included.

• Epsilon (ε): The privacy budget quantifying maximum information leakage
• Laplace Mechanism: Adds noise scaled to query sensitivity for numeric outputs
• Composability: Tracks cumulative privacy loss across multiple queries

A privacy model ensuring each record is indistinguishable from at least k-1 other records with respect to quasi-identifiers. It prevents singling out individuals but requires careful handling of homogeneity attacks.

• Generalization: Replacing specific values with broader categories (e.g., age 34 → 30-40)
• Suppression: Removing high-risk attributes or records entirely
• Quasi-identifier (QID): Non-sensitive attributes linkable to external data

A cryptographic scheme enabling computation directly on ciphertexts. It allows ML inference on encrypted data without exposing raw inputs to the model host, enforcing PbD's 'full functionality' principle.

• Partially Homomorphic (PHE): Supports only addition or multiplication
• Fully Homomorphic (FHE): Supports arbitrary circuits with high computational cost
• Leveled FHE: Practical trade-off limiting circuit depth for performance

A protocol allowing multiple parties to jointly compute a function over their private inputs without revealing those inputs to each other. It operationalizes PbD by distributing trust.

• Garbled Circuits: Boolean circuit representation evaluated obliviously
• Secret Sharing: Splitting data into meaningless shares distributed across parties
• Input Privacy: Each party learns only the final output, nothing else

The processing of personal data to replace direct identifiers with artificial pseudonyms. Unlike anonymization, re-identification remains possible with separately stored additional information.

• Tokenization: Substituting sensitive fields with non-sensitive surrogates
• Keyed Hash: Using a secret key to generate deterministic pseudonyms
• Art. 4(5) GDPR: Explicitly defined as a privacy-enhancing safeguard