14. Security & Privacy

Data Encryption

In transit: TLS 1.3 for all inter-service and client-service communication.
At rest: AES-256 encryption for storage (S3, database, feature store).
Key management: Use KMS (AWS KMS, Azure Key Vault, Vault) for key rotation and access control.

Role-Based Access Control (RBAC)

Role	Access to Data	Access to Models	Access to Infra
Data Engineer	Raw data, feature store (R/W)	—	Data pipelines
ML Engineer	Feature store (R), training data	Train, register, deploy	Training, serving
MLOps	Monitoring (R)	Deploy, rollback	All infra
Data Scientist	Feature store (R), predictions (R)	View registry	—
Auditor	Logs (R)	Registry metadata (R)	—

Differential Privacy

Adds calibrated noise to training data or model updates to protect individual privacy.

# Differential privacy during gradient descent
epsilon = 8.0   # privacy budget
delta = 1e-5    # failure probability

for batch in dataloader:
    gradients = compute_gradients(batch, model)
    # Clip gradients to limit individual influence
    gradients = clip(gradients, max_norm=1.0)
    # Add noise proportional to epsilon
    noise = gaussian_noise(scale=1.0/epsilon)
    gradients += noise
    optimizer.apply(gradients)

Bias & Fairness

Pre-training: Audit training data for representation bias.
Post-training: Evaluate model performance across demographic groups.
Post-deployment: Monitor prediction rates by group; alert on disparity.

# Fairness monitoring per group
Group          | Approval Rate | False Positive | False Negative
Male (50K)     | 72%           | 3.2%           | 2.1%
Female (45K)   | 71%           | 3.4%           | 1.9%
Other (5K)     | 65%           | 5.1%           | 3.8%  ⚠ Disparity

Action: Investigate data representation for "Other" group

Compliance (GDPR, CCPA, HIPAA)

Right to be forgotten: Delete user data from feature store + retrain model without that user's data.
Data portability: Export user's prediction history.
Explainability: Store SHAP values with each prediction for audit.
Consent tracking: Tag each data point with consent level; respect in feature computation.
Data retention: Automatic deletion of raw data after retention period.

Exercise: You are building an ML system for a healthcare provider (HIPAA-regulated) that predicts patient readmission risk. Design the security and privacy architecture: a) encryption strategy, b) RBAC roles with read/write permissions, c) differential privacy considerations, d) fairness auditing plan (which groups to monitor), e) how to handle a patient's request to delete their data.