Anonymization

A mathematical framework providing a provable guarantee that the output of a computation reveals no information about any single individual's presence in the dataset. Controlled by the privacy loss parameter epsilon (ε).

• Laplace Mechanism: Adds noise calibrated to L1 sensitivity for pure ε-differential privacy
• Gaussian Mechanism: Adds noise calibrated to L2 sensitivity for (ε, δ)-differential privacy
• Privacy Budget: A finite resource tracking cumulative ε expenditure across queries
• Post-processing Immunity: Any computation on a differentially private output remains differentially private

A data de-identification technique that replaces direct identifiers (names, medical record numbers) with artificial pseudonyms. Unlike anonymization, pseudonymized data can be re-linked to an individual using separately held key information.

• Retains analytical utility while reducing re-identification risk
• Still considered personal data under GDPR
• Common in clinical research where longitudinal tracking is required
• Example: Replacing 'John Smith, MRN 12345' with 'Subject A7X2'

A privacy property ensuring each released record is indistinguishable from at least k-1 other records with respect to quasi-identifiers (age, ZIP code, gender). Prevents re-identification through linking attacks with external datasets.

• Achieved through generalization (replacing specific values with ranges) and suppression (removing outlier records)
• Vulnerable to homogeneity attacks when sensitive attributes lack diversity within k-groups
• Example: Releasing only 5-digit ZIP codes instead of 9-digit, and age ranges instead of exact ages

A mathematical framework that provides a provable privacy guarantee by adding calibrated noise to query outputs. The epsilon (ε) parameter quantifies the privacy loss—a smaller epsilon means stronger protection.

• Ensures an adversary cannot determine whether any single individual's data was included
• Uses mechanisms like Laplace (pure ε-DP) and Gaussian ((ε, δ)-DP)
• Deployed by the US Census Bureau for 2020 Census data releases
• Trade-off: Higher privacy (lower ε) reduces statistical accuracy

A cryptographic primitive enabling computation directly on encrypted data without decryption. The result, when decrypted, matches what would have been obtained by operating on the plaintext.

• Partially Homomorphic Encryption (PHE): Supports only addition OR multiplication
• Somewhat Homomorphic Encryption (SHE): Supports limited operations
• Fully Homomorphic Encryption (FHE): Supports arbitrary computation but remains computationally intensive
• Critical for cloud-based analysis of encrypted medical records

A cryptographic protocol allowing multiple parties to jointly compute a function over their private inputs while revealing nothing beyond the designated output. No party learns another's raw data.

• Enables cross-institutional research without sharing patient-level data
• Uses techniques like secret sharing and garbled circuits
• Example: Three hospitals compute average patient outcome without revealing individual hospital statistics
• Provides input privacy even against honest-but-curious adversaries

An adversarial attack that reconstructs representative training data from access to a trained model's parameters or predictions. Demonstrates why anonymization alone is insufficient for model protection.

• Can recover facial images from facial recognition models
• Exploits confidence scores or gradient information
• Mitigation requires differential privacy during training or limiting output granularity
• Highlights that trained models memorize aspects of their training data