Title: AWS re:Inforce 2024 - Secure your healthcare generative AI workloads on Amazon EKS (DAP221)

Insights:

• Introduction and Purpose: The session aims to educate on hardening container images, minimizing attack surfaces, and implementing secure container orchestration for generative AI workloads in healthcare using Amazon EKS.
• Generative AI Use Cases in Healthcare:
  • AWS HealthScribe: Automatically creates clinical notes from patient-clinician conversations.
  • SageMaker: Used for jumpstarting foundational models or fine-tuning pre-trained models from Hugging Face.
  • Amazon Pharmacy: Utilized for Q&A chatbots using retrieval-augmented generation.
  • Open Repositories: Cancer Imaging Archive and Imaging Data Commons democratize access to large imaging data for machine learning analysis.
• Threat Landscape Using STRIDE Framework:
  • Spoofing: False authenticity leading to model hijacking and malicious use.
  • Tampering: Introduction of malicious data into AI model training, causing unintended behaviors.
  • Repudiation: False claims and manipulations, often leading to model stealing or side-channel attacks.
  • Information Disclosure: Data leaks and model extraction attacks compromising confidentiality.
  • Denial of Service: Resource exhaustion and adversarial attacks disrupting healthcare operations.
  • Elevation of Privileges: Authorization leakage requiring runtime controls to prevent damage to reputation and legal issues.
• Detailed Walkthrough:
  • Kubernetes and EKS: Celebrating 10 years, Kubernetes offers vast potential for generative AI workloads.
  • Intelligent Health Assistant Use Case:
    • Biomistril-7b Model: A large language model trained with PubMed central data, supporting multiple languages.
    • EKS Deployment: Core node group for Kubernetes cluster utilities, just-in-time node scaler (Carpenter), and inferential nodes for ML workloads.
    • API Exposure: Fast API Python app with network load balancer or web UI tool like Gradio for user interaction.
• Security Measures and Best Practices:
  • Spoofing Mitigation: Least privileged access using role-based access control and IAM, with audit logs via CloudTrail.
  • Tampering Mitigation: Ensuring integrity with policy as code solutions, cryptographic signing, and vulnerability scanning.
  • Repudiation Mitigation: Strong telemetry and monitoring with logging drivers and analytics platforms like OpenSearch or CloudWatch.
  • Information Disclosure Mitigation: Encrypting secrets with Key Management Service and Secrets Manager.
  • Denial of Service Mitigation: Preventative measures like network policies, sub-segmentation, and tools like AWS Shield and Firewall Manager.
  • Elevation of Privileges Mitigation: Resource limits, runtime tools like GuardDuty, and automated remediation with AWS Config.
• Key Takeaways:
  • Generative AI use cases and vulnerability landscape.
  • STRIDE framework application.
  • Architectural patterns and AWS security best practices.
  • Reference to the Data on EKS program for well-architected blueprints.

Quotes:

• "Are you interested in learning techniques for hardening container images, minimizing attack surfaces, and implementing secure container orchestration?"
• "94% of executives say it's important to secure AI solutions before deployment, which is why we're talking to you today."
• "The implications for spoofing include model hijacking, where attackers could gain control of the generative AI model and use it for malicious purposes."
• "Tampering refers to the intentional introduction of malicious data into the training process of the AI model."
• "Generative AI models can be vulnerable to data leaks and model extraction attacks."
• "Denial of service is where resource exhaustion is experienced by denying service, and it usually happens at the network layer."
• "Elevation of privileges entails authorization leakage and requires runtime controls."
• "Kubernetes is celebrating 10 years, just four days ago, which is a good thing."
• "The best part is it's been trained with the PubMed central data, benchmarked across ten different general Q&A, and is multilingual."
• "Please do reach out to your account teams to engage us and know more about what more than what we could cover in this 20 minutes."