Intrigued by the art of cloud architecture? Discover how to design, develop, and manage robust, secure, scalable, and dynamic solutions on Google Cloud as you prepare for the Professional Cloud Architect exam!
Gauge your current knowledge
Gauge your current knowledge
Google Cloud automatically encrypts data at rest, but for more control, you can use Customer-Managed Encryption Keys (CMEK). These keys are managed through Cloud Key Management Service (Cloud KMS), letting you control the key's rotation schedule and who can access it. Using CMEK protects sensitive AI assets like training datasets, model artifacts, and batch prediction results according to your own security standards.
Managing who can use these resources is done through Identity and Access Management (IAM). You must grant specific service accounts the Cloud KMS CryptoKey Encrypter/Decrypter role, following the principle of least privilege. This ensures only approved services can encrypt or decrypt your AI data. Using audit logging provides a record of when and how these keys are used, giving you visibility and control.
To stop data from being stolen or leaked, use VPC Service Controls (VPC-SC) to create a service perimeter. This acts like a digital fence around your AI projects, blocking unauthorized data transfers out of your trusted network. A common unified perimeter is often the best design because it provides strong protection against data exfiltration while keeping the setup simple.
For the most sensitive workloads, Confidential Space provides a Trusted Execution Environment (TEE). This keeps data private even from Google's system operators while the AI model runs. It uses a hardened OS image to prevent tampering. These tools, along with services for detecting sensitive data, create a highly secure foundation for deploying AI.
Sensitive Data Protection is a service used to find and manage Personally Identifiable Information (PII), like credit card numbers or addresses, within your datasets. Using automated scanning, you can discover this sensitive information across large datasets before using the data for AI training or inference. This is the first step to ensuring the data is handled safely and in compliance with regulations.
De-identification is the process of removing or hiding sensitive identifiers while keeping the data useful for analysis. You apply consistent rules using transformation templates. Common techniques include masking (hiding parts of a value), tokenization (replacing data with a safe token), and bucketing (grouping values into ranges). This protects individual privacy while still allowing the AI model to learn patterns.
Data sovereignty and data residency are important for meeting global privacy laws. You can use Customer-Managed Encryption Keys (CMEK) to manage your own encryption keys, rather than using Google's defaults. You can also set resource location policies to ensure your AI data stays within specific geographic regions, complying with standards like HIPAA.
Advanced tools add extra security layers. Differential Privacy works by adding statistical noise to data, making it very hard to identify any single person while still allowing the model to learn. Monitoring through Cloud Audit Logs creates a clear record of who accessed sensitive data and when, providing a trail for security reviews.
Model Armor is a primary tool for protecting Large Language Models (LLMs) from attacks like prompt injections and jailbreaking attempts. It screens both what users ask and what the model answers to block prohibited or harmful content. By adding this security layer early, you prevent unsafe outputs from reaching users.
Model Armor integrates with platforms like Vertex AI and Google Kubernetes Engine (GKE). In GKE, it uses Service Extensions to monitor traffic to and from the model. Secure model deployment also uses Binary Authorization to ensure only verified and trusted container images can run in production, preventing malicious code from being deployed.
Administrators can configure safety filters and content moderation thresholds to balance the model's usefulness with security needs. These settings let you block harmful, biased, or unwanted outputs. Key configuration tools include floor settings for global safety rules, templates for standardizing security, and sanitization to clean inputs and outputs and prevent sensitive data disclosure.
Continuous monitoring is essential. Tools like Cloud Logging and the AI Security Dashboard give visibility into sanitization results and help track security events. Sensitive Data Protection can also be used to find and mask PII within the AI's prompts. Finally, enforcing the principle of least privilege via IAM ensures only authorized users can change model settings or access sensitive training data.
StudyCerts
StudyCerts