33 lines (21 loc) · 1.19 KB

SECURITY.md – ML-Based Risk Scoring – Tier-1 UK Retail Bank (GCP + BigQuery ML)

This document describes security and governance patterns for the case study.

1. Perimeter and network

All data and services reside inside a VPC-SC protected perimeter.
Access from outside is via approved service accounts and identities only.

2. Encryption

All BigQuery tables (raw, features, scores) use CMEK.
GCS buckets for model artifacts and DQ exports use CMEK as well.
TLS in transit is enforced by GCP-managed certificates.

3. Identity & access management

IAM follows least-privilege:
- Dataflow, Composer, and BQML have separate service accounts.
- Read vs write vs admin roles split by team.
Row-level and column-level security used for sensitive attributes.

4. Data classification and policy tags

Sensitive columns (PII, financial identifiers) are tagged with Policy Tags.
Access to tagged columns is restricted to specific roles.

5. Audit logging

Cloud Audit Logs enabled for BigQuery, Dataflow, Pub/Sub, Composer.
Access to risk scores and model metadata is logged and retained.

This is a sanitized description designed for portfolio purposes only.