Securing AI Systems with AWS Services

​

Identity and Access Management (IAM)

• Manage users, groups, roles, policies, and federation with IAM.
• Enforce Multi-Factor Authentication (MFA) for privileged and console users.
• Apply the principle of least privilege: grant the minimum permissions required.
• Use IAM roles and temporary credentials for applications, services, and cross-account access.

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "s3:ListBucket"
      ],
      "Resource": [
        "arn:aws:s3:::example-bucket"
      ]
    },
    {
      "Effect": "Allow",
      "Action": [
        "s3:GetObject",
        "s3:PutObject"
      ],
      "Resource": [
        "arn:aws:s3:::example-bucket/*"
      ]
    }
  ]
}

​

Root account and user management

• Avoid using the root AWS account for routine tasks—use it only for rare account-level changes (billing, root-level contact info, or closing the account).
• Create IAM users and place them into groups with role-appropriate permissions.
• Prefer group-based permission assignment for easier auditing and maintenance.

​

Roles and temporary credentials

• IAM roles bundle permissions with a trust policy; when assumed, they provide temporary credentials that expire.
• Use roles for EC2 instance profiles, Lambda execution roles, cross-account access, and federation with identity providers.

​

Logging and auditing

• Enable AWS CloudTrail to record API activity across your account for auditing and incident investigation.
• Store CloudTrail logs in a dedicated, secured S3 bucket in a separate logging or security account to prevent tampering.
• Apply S3 public access blocking and strict bucket policies for log storage and retention.

​

SageMaker role separation

• Create distinct roles for data scientists, MLOps engineers, and compute/service accounts.
• Limit dataset access and model artifact permissions according to job needs.

​

Encryption

• AWS Key Management Service (KMS) manages cryptographic keys for server-side encryption.
• S3 supports server-side encryption options: SSE-S3, SSE-KMS, and customer-provided keys (SSE-C).
• For highly sensitive data, use client-side encryption so data is encrypted before upload.
• Use TLS/SSL for data in transit and store certificates in AWS Certificate Manager (ACM) for integration with load balancers, API Gateway, and CloudFront.

• Use ACM to provision and manage TLS certificates for AWS-integrated services.
• AWS SDKs and the CLI default to HTTPS; ensure certificates are valid and renew before expiry.

​

Network isolation and private connectivity

• A VPC provides an isolated virtual network spanning Availability Zones (AZs). Use public subnets for internet-facing resources and private subnets for internal components.
• Private subnets that need outbound internet access should use a NAT gateway or NAT instance. To avoid routing traffic over the internet to reach AWS public services, use VPC endpoints.
• Gateway endpoints are available for S3 and DynamoDB; interface endpoints (AWS PrivateLink) provide private connectivity to many other AWS services.
• Apply security groups (stateful) and network ACLs (stateless) for layer-4 access control. Consider AWS Network Firewall or third-party appliances for additional inspection.

​

Inter-node encryption for distributed training

• Enable SageMaker inter-node encryption for distributed training jobs to encrypt network traffic between training instances. This is particularly important when workers exchange sensitive features or PII.

​

VPCs, subnets, and internet access

• Multi-AZ VPC designs typically include public and private subnets for high availability and separation of concerns.
• Public subnets route through an Internet Gateway; private subnets use a NAT gateway for outbound internet access while remaining unreachable from the internet.
• Combine security groups, NACLs, and endpoint policies to minimize exposure.

​

VPC endpoints

• Use gateway endpoints for S3 and DynamoDB.
• Use interface endpoints (PrivateLink) for other supported services to keep traffic on the AWS network and reduce exposure to the public internet.

​

Quick reference: Common AWS security services and use cases

​

Summary

• Shared Responsibility Model — what AWS manages vs. what you must manage.
• IAM best practices — policies, roles, MFA, and least privilege.
• Logging and auditing — CloudTrail and secure log storage.
• Encryption — KMS for keys, server-side vs. client-side encryption, and ACM for TLS.
• Network isolation — VPC design, subnets, NAT, VPC endpoints, and PrivateLink.
• SageMaker-specific considerations — role separation and inter-node encryption for distributed training.

• AWS Security Documentation
• AWS KMS Developer Guide
• Amazon SageMaker Security