Skip to main content
Encryption is a core control for protecting sensitive machine learning (ML) datasets. It helps secure personally identifiable information (PII) and protected health information (PHI) across financial and healthcare domains, ensures data integrity, prevents unauthorized modification, and supports regulatory compliance (for example, GDPR, HIPAA, and PCI DSS). Apply encryption across three stages of the data lifecycle:
  • At rest — protect stored data (for example, Amazon S3 objects, EBS volumes, RDS storage, and model artifacts).
  • In transit — protect data moving between clients, services, and API endpoints.
  • During processing — protect data in use (for example, encrypted temporary storage, VPC isolation, or enclave technologies).
Each stage requires different controls and trade-offs between security and usability. The sections below explain practical AWS controls, key-management considerations, and recommended patterns for ML workloads.

​
Encryption across the ML data lifecycle (summary)

​
Encryption at rest

Protect stored datasets, feature stores, model artifacts, and backups using storage-level encryption. Common AWS services and options include:
  • Amazon S3: Server-Side Encryption (SSE) encrypts objects after upload and decrypts on authorized access. S3 supports:
  • Amazon EBS: Use AWS KMS-backed encryption for volumes and snapshots. Encryption is transparent to the instance and applies to data at rest and data moving between the instance and volume. Enable encryption at volume creation or create an encrypted copy of an unencrypted snapshot. Docs: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/EBSEncryption.html
  • Amazon RDS: RDS supports storage encryption using AWS KMS for supported engines (MySQL, MariaDB, PostgreSQL, Oracle, SQL Server, and Amazon Aurora). Encryption applies to underlying storage, automated backups, read replicas, and snapshots. Encryption must be enabled at instance creation. See RDS docs: https://learn.kodekloud.com/user/courses/aws-rds
The image illustrates AWS options for encryption at rest, showing a flow from Amazon EC2 to Amazon EBS, with AWS KMS for key management.

​
Practical examples

Enable S3 default encryption for a bucket:
Create an encrypted EBS volume (example snippet):
Note: For existing unencrypted RDS or EBS resources, create snapshots and restore encrypted copies rather than attempting in-place conversion.

​
Key management

Centralized, auditable key management is essential for secure ML pipelines.
  • AWS KMS provides managed keys (AWS-managed) and customer-managed CMKs for finer-grained control.
  • Use CMKs when you need to apply specific IAM and key policies, enable automatic rotation, or audit usage through CloudTrail.
  • Integrations: When services like S3, EBS, and RDS use KMS, the services call KMS on behalf of the service principal while enforcing IAM and KMS policies.
The image is a diagram explaining "Encryption at Rest" options in AWS, specifically using Amazon RDS with AWS Key Management Service (KMS) for databases like MySQL, MariaDB, and PostgreSQL.

​
S3, EBS, and RDS — quick recap

Next: When integrating S3 with KMS, prefer SSE-KMS with a customer-managed key to gain control over usage policies and CloudTrail visibility.
The image illustrates AWS Key Management Service (AWS KMS), showing a user interacting with both Customer Managed Keys (CMK) and AWS Managed Keys for services like SQS, S3, and EBS.

​
Encryption in transit

Encryption in transit prevents interception and tampering while data moves between clients and services.
  • Use TLS (HTTPS) for all client-to-service interactions (S3 API calls, SageMaker endpoints, model inference APIs).
  • For intra-VPC communication, prefer VPC endpoints and private networking so traffic does not traverse the public internet.
  • Where possible, enforce TLS for service-to-service calls and enable mutual TLS for higher assurance.

​
Client-side encryption

Client-side encryption encrypts data before upload so that only ciphertext is stored in S3. It is suitable where you must ensure cloud services never see plaintext.
  • Pros: Strong privacy guarantees and data-control assurances.
  • Cons: Added complexity for key distribution, rotation, and multi-user access.
  • Use established libraries such as the AWS Encryption SDK or well-vetted cryptographic libraries. Avoid custom crypto implementations.
The image illustrates client-side encryption using Amazon S3, depicting a user encrypting data before storing it in Amazon's Simple Storage Service.

​
Encryption during ML processing

Protect data in use by limiting plaintext exposure during training and inference:
  • SageMaker and other managed ML services support encryption for input datasets, model artifacts, and TLS for endpoints.
  • Use VPC isolation, VPC endpoints, and KMS-backed encryption for artifacts and training data to minimize exposure.
  • For high-assurance workloads, consider dedicated secure compute options:
    • Nitro Enclaves for isolated, attested compute environments
    • Dedicated instances or dedicated hosts
    • Workload-level isolation (separate training jobs, separate roles, or separate accounts)
The image is a diagram illustrating encryption in AWS SageMaker, showing components like VPC, private subnets, and various AWS services interconnected through an elastic network interface and VPC endpoints.

​
Best practices and recommendations

  • Encryption scope: Ensure you cover raw data, feature stores, model artifacts, logs, and backups.
  • Key management: Use AWS KMS for central key management, rotation, and auditing. Prefer CMKs for granular policies, AWS-managed keys for simplicity.
  • Service-level encryption: Enable SSE-KMS for S3, EBS encryption, RDS encryption, and encrypt Feature Store and SageMaker artifacts. See SageMaker Feature Store: https://docs.aws.amazon.com/sagemaker/latest/dg/feature-store.html
  • Least privilege: Apply IAM and KMS key policies so only authorized roles can decrypt or use keys.
  • Monitoring and auditing: Enable CloudTrail logging for KMS and monitor key usage for anomalies.
  • Data lifecycle: Plan secure backups, handle snapshots safely, and define secure deletion policies (for example, handle KMS key deletion carefully and manage snapshot lifecycles).
Enable encryption at resource creation (for example, RDS and EBS). For existing unencrypted resources, create encrypted copies (snapshots or restores) rather than trying to flip encryption on the original resource in-place.
The image outlines best practices for ML data encryption, including encrypting data at rest and in transit, using AWS KMS for key management, and enabling encryption for various AWS services.

​
Additional resources

Adopting these encryption controls across the ML data lifecycle reduces the risk of data exposure, strengthens compliance posture, and helps secure model development and deployment workflows.

Watch Video