How do you handle PII and data privacy across an ML pipeline on AWS?

Detect and classify PII with Macie or Comprehend, minimise and tokenise at ingestion, enforce column-level access with Lake Formation, encrypt with scoped KMS keys, and design for deletion requests by keeping identifiers separable from features.

Privacy in ML is harder than in ordinary applications because data is copied into training sets, embedded into model weights, and cached in feature stores — all places a naive deletion process misses.

1. Discover and classify. You cannot protect what you have not found. Amazon Macie scans S3 for PII; Comprehend detects it in free text at ingestion:

```python comprehend = boto3.client("comprehend")…