What's changed: In-scope coverage: monitoring/ops/cost/security services
4.2ML Security and Cost
Understand security/cost of ML operations: least privilege with IAM, encryption (KMS/TLS) and VPC isolation, and cost optimization (right-sizing, managed spot training).
Like any AWS workload, ML solutions are protected with least privilege, encryption, and isolation, and their training/inference cost is optimized.
4.2.1Security and cost
- IAM least privilege: give training jobs/endpoints roles with minimal permissions (no embedded keys).
- Encryption & isolation: encrypt data with KMS (at rest)/TLS (in transit); isolate networking with a VPC for private communication.
- Cost optimization: right-size instances, use managed spot training, and save on inference with serverless/multi-model.
Common on MLA: cut training cost = managed spot training (with checkpoints for interruptions), data protection = KMS/TLS + VPC isolation, least-privilege IAM roles for jobs/endpoints.
ML operations are secured like any AWS workload—"least privilege, encryption, isolation"—while optimizing training/inference cost. IAM least privilege attaches an execution role to training jobs/endpoints, scoped to only the needed S3 prefixes and KMS keys (no embedded long-lived keys). Encryption uses KMS at rest (S3/EBS/model artifacts/volumes) and TLS in transit. Isolate jobs/endpoints in a VPC, reach S3/SageMaker APIs without the internet via VPC endpoints (PrivateLink), and optionally use network isolation mode (block container egress). Audit with CloudTrail and find sensitive data with Macie. Optimize cost with "training = managed spot training (up to ~90% off, with checkpoints for interruptions)" and "inference = traffic-based Auto Scaling, serverless for intermittent, multi-model endpoints for many models," using Inference Recommender to right-size instances. The axes: "training cost = spot + checkpoints," "protection = least privilege + KMS/TLS + VPC."
| Goal | Means |
|---|---|
| Job/endpoint permissions | Least-privilege execution role |
| Data protection | KMS (rest) / TLS (transit) |
| Network isolation | VPC + VPC endpoints |
| Cut training cost | Managed spot + checkpoints |
Scenario: train cheaply on sensitive data and serve inference only from the internal network. Train with managed spot + S3 checkpoints to cut cost. Encrypt data with KMS and run jobs under a least-privilege execution role. Place the endpoint in a VPC with VPC endpoints (PrivateLink) for private traffic, optionally with network isolation mode. Audit actions with CloudTrail.
Q. Cheaper training? Managed spot + checkpoints. Q. Job S3 access? Least-privilege execution role (no embedded keys). Q. Data protection? KMS/TLS. Q. Network isolation? VPC + VPC endpoints. Q. Inference cost? Auto Scaling/serverless/multi-model.
Watch the mix-ups: (1) Managed spot can be interrupted—checkpoints are essential (don’t use it for low-latency production inference). (2) Never embed long-lived keys—use execution roles. (3) VPC-bound jobs need VPC endpoints to reach S3/SageMaker APIs (otherwise unreachable). (4) Encryption still needs separate access control (encryption ≠ authorization).
Managed spot training can be interrupted, so checkpoint to S3 to resume; for inference, auto-scale with traffic to control cost.
4.2.2In-scope services for monitoring, operations, cost, and security
For monitoring and maintaining ML solutions, collect/analyze app logs with Amazon CloudWatch Logs, do ML anomaly detection on operational data with Amazon DevOps Guru, and audit API operations with CloudTrail. Manage fleet config, patching, and parameters/secrets with AWS Systems Manager, and store secrets encrypted with auto-rotation via AWS Secrets Manager. Right-size resources with AWS Compute Optimizer, check best practices with AWS Trusted Advisor, and relay alerts to Slack/Teams with AWS Chatbot. For cost, the umbrella for billing and cost management is AWS Billing and Cost Management, visualization/forecast is AWS Cost Explorer, and budget-overrun alerts are AWS Budgets. Govern multi-account ML environments (SCPs/consolidated billing) with AWS Organizations.
4.2.3Section summary
- Protect with least-privilege IAM / KMS·TLS + VPC isolation
- Training cost = managed spot + checkpoints
- Monitor = CloudWatch Logs/DevOps Guru; ops = Systems Manager/Secrets Manager; cost = Cost Explorer/Budgets; govern = Organizations
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. You want to train much cheaper while tolerating interruptions. Which fits best?
Q2. What is the secure way for a training job to access data in S3?
Q3. You want to isolate inference endpoints/data on the network for private communication. What do you use?
Keep track of your progress
The full study guide is free to read. Sign up free to practice with the question bank, track what you have read, review your mistakes, and highlight passages.