Editing Privacy-Preserving Machine Learning (section)

== <span style="color: #FFFFFF;">Remembering</span> ==
* '''Differential privacy (DP)''' — A mathematical guarantee that the inclusion or exclusion of any single record makes little difference to the output, bounded by parameter ε.
* '''Epsilon (ε) in DP''' — The privacy budget: smaller ε = stronger privacy guarantee but more noise added. Typical values: ε=1–10.
* '''Noise mechanism''' — Adding calibrated random noise to protect privacy: Laplace mechanism, Gaussian mechanism.
* '''DP-SGD (Differentially Private SGD)''' — Training neural networks with differential privacy by clipping and noising gradients.
* '''Federated learning''' — Training on data distributed across many devices without centralizing the raw data; only model updates are shared.
* '''Secure aggregation''' — Aggregating federated model updates without the server seeing individual updates (using cryptographic protocols).
* '''Homomorphic encryption (HE)''' — Cryptographic technique allowing computation on encrypted data without decryption.
* '''Secure Multi-Party Computation (SMPC)''' — Multiple parties jointly compute a function on their private inputs without revealing those inputs.
* '''Membership inference attack''' — An attack testing whether a specific record was in the training data; measures privacy leakage.
* '''Model inversion attack''' — Reconstructing training data from a trained model's outputs; a privacy risk.
* '''Data minimization''' — Collecting and using only the minimum data necessary; a GDPR principle.
* '''Synthetic data (privacy)''' — Generating realistic but non-personal data to share instead of real records.
* '''k-anonymity''' — A data protection model where each record is indistinguishable from at least k-1 others.
* '''Privacy budget''' — The total privacy expenditure across multiple DP queries or training steps; must be managed carefully.
</div>

<div style="background-color: #006400; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">