This principle in the security pillar of the [Google Cloud Well-Architected Framework](https://docs.cloud.google.com/architecture/framework) provides recommendations to help you secure your AI systems. These recommendations are aligned with Google's [Secure AI Framework (SAIF)](https://safety.google/cybersecurity-advancements/saif/), which provides a practical approach to address the security and risk concerns of AI systems. SAIF is a conceptual framework that aims to provide industry-wide standards for building and deploying AI responsibly. ## Principle overview To help ensure that your AI systems meet your security, privacy, and compliance requirements, you must adopt a holistic strategy that starts with the initial design and extends to deployment and operations. You can implement this holistic strategy by applying the [six core elements of SAIF](https://developers.google.com/machine-learning/resources/saif). Google uses AI to enhance security measures, such as identifying threats, automating security tasks, and improving detection capabilities, while keeping humans in the loop for critical decisions. Google emphasizes a collaborative approach to advancing AI security. This approach involves partnering with customers, industries, and governments to enhance the SAIF guidelines and offer practical, actionable resources. The recommendations to implement this principle are grouped within the following sections: - [Recommendations to use AI securely](https://docs.cloud.google.com/architecture/framework/security/use-ai-securely-and-responsibly#recommendations_to_use_ai_securely) - [Recommendations for AI governance](https://docs.cloud.google.com/architecture/framework/security/use-ai-securely-and-responsibly#recommendations_for_ai_governance) ## Recommendations to use AI securely To use AI securely, you need both foundational security controls and AI-specific security controls. This section provides an overview of recommendations to ensure that your AI and ML deployments meet the security, privacy, and compliance requirements of your organization. For an overview of architectural principles and recommendations that are specific to AI and ML workloads in Google Cloud, see the [AI and ML perspective](https://docs.cloud.google.com/architecture/framework/perspectives/ai-ml) in the Well-Architected Framework. ### Define clear goals and requirements for AI usage This recommendation is relevant to the following [focus areas](https://docs.cloud.google.com/architecture/framework/security#focus_areas_of_cloud_security): - Cloud governance, risk, and compliance - AI and ML security This recommendation aligns with the SAIF element about contextualizing AI system risks in the surrounding business processes. When you design and evolve AI systems, it's important to understand your specific business goals, risks, and compliance requirements. ### Keep data secure and prevent loss or mishandling This recommendation is relevant to the following [focus areas](https://docs.cloud.google.com/architecture/framework/security#focus_areas_of_cloud_security): - Infrastructure security - Identity and access management - Data security - Application security - AI and ML security This recommendation aligns with the following SAIF elements: - Expand strong security foundations to the AI ecosystem. This element includes data collection, storage, access control, and protection against data poisoning. - Contextualize AI system risks. Emphasize data security to support business objectives and compliance. ### Keep AI pipelines secure and robust against tampering This recommendation is relevant to the following [focus areas](https://docs.cloud.google.com/architecture/framework/security#focus_areas_of_cloud_security): - Infrastructure security - Identity and access management - Data security - Application security - AI and ML security This recommendation aligns with the following SAIF elements: - Expand strong security foundations to the AI ecosystem. As a key element of establishing a secure AI system, secure your code and model artifacts. - Adapt controls for faster feedback loops. Because it's important for mitigation and incident response, track your assets and pipeline runs. ### Deploy apps on secure systems using secure tools and artifacts This recommendation is relevant to the following [focus areas](https://docs.cloud.google.com/architecture/framework/security#focus_areas_of_cloud_security): - Infrastructure security - Identity and access management - Data security - Application security - AI and ML security Using secure systems and validated tools and artifacts in AI-based applications aligns with the SAIF element about expanding strong security foundations to the AI ecosystem and supply chain. This recommendation can be addressed through the following steps: - Implement a secure environment for ML training and deployment - Use validated container images - Apply [Supply-chain Levels for Software Artifacts (SLSA)](https://slsa.dev) guidelines ### Protect and monitor inputs This recommendation is relevant to the following [focus areas](https://docs.cloud.google.com/architecture/framework/security#focus_areas_of_cloud_security): - Logging, auditing, and monitoring - Security operations - AI and ML security This recommendation aligns with the SAIF element about extending detection and response to bring AI into an organization's threat universe. To prevent issues, it's critical to manage prompts for generative AI systems, monitor inputs, and control user access. ## Recommendations for AI governance All of the recommendations in this section are relevant to the following [focus area](https://docs.cloud.google.com/architecture/framework/security#focus_areas_of_cloud_security): Cloud governance, risk, and compliance. Google Cloud offers a robust set of tools and services that you can use to build responsible and ethical AI systems. We also offer a framework of policies, procedures, and ethical considerations that can guide the development, deployment, and use of AI systems. As reflected in our recommendations, Google's approach for AI governance is guided by the following principles: - Fairness - Transparency - Accountability - Privacy - Security ### Use fairness indicators [Gemini Enterprise Agent Platform](https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning) can detect bias during the data collection or post-training evaluation process. Agent Platform provides [model evaluation metrics](https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/evaluation/intro-evaluation-fairness) like *data bias* and *model bias* to help you evaluate your model for bias. These metrics are related to fairness across different categories like race, gender, and class. However, interpreting statistical deviations isn't a straightforward exercise, because differences across categories might not be a result of bias or a signal of harm. ### Use Vertex Explainable AI To understand how the AI models make decisions, use Vertex Explainable AI. This feature helps you to identify potential biases that might be hidden in the model's logic. This explainability feature is integrated with [BigQuery ML](https://docs.cloud.google.com/bigquery/docs/xai-overview) and [Vertex Explainable AI](https://docs.cloud.google.com/vertex-ai/docs/explainable-ai/overview), which provide feature-based explanations. You can either perform explainability in BigQuery ML or [register your model](https://docs.cloud.google.com/bigquery/docs/managing-models-vertex#register_models) in Agent Platform and perform explainability in Agent Platform. ### Track data lineage Track the origin and transformation of data that's used in your AI systems. This tracking helps you understand the data's journey and identify potential sources of bias or error. [Data lineage](https://docs.cloud.google.com/data-catalog/docs/concepts/about-data-lineage) is a Knowledge Catalog feature that lets you track how data moves through your systems: where it comes from, where it's passed to, and what transformations are applied to it. ### Establish accountability Establish clear responsibility for the development, deployment, and outcomes of your AI systems. Use [Cloud Logging](https://docs.cloud.google.com/logging/docs/overview) to log key events and decisions made by your AI systems. The logs provide an audit trail to help you understand how the system is performing and identify areas for improvement. Use [Error Reporting](https://docs.cloud.google.com/error-reporting/docs/grouping-errors) to systematically analyze errors made by the AI systems. This analysis can reveal patterns that point to underlying biases or areas where the model needs further refinement. ### Implement differential privacy During model training, [add noise](https://docs.cloud.google.com/bigquery/docs/differential-privacy#add_noise) to the data in order to make it difficult to identify individual data points but still enable the model to learn effectively. With [SQL in BigQuery](https://docs.cloud.google.com/bigquery/docs/introduction-sql), you can transform the results of a query with differentially private [aggregations](https://docs.cloud.google.com/bigquery/docs/differential-privacy).