Google Launches Framework to Secure Generative AI

Generative AI is advancing rapidly, but so are the creative ways to use it maliciously. Many governments are accelerating their regulatory plans to mitigate the risks of AI abuse.

Meanwhile, some generative AI developers are looking at ways to secure their models and services. Google, owner of generative AI chatbot Bard and parent company of AI research lab DeepMind, introduced the Secure AI Framework (SAIF) on June 8, 2023.

SAIF stands for “Bold, Responsible, […] It’s a conceptual framework for jointly securing the security of AI technology,” said Royal Hansen, Google’s vice president of engineering for privacy, safety and security, and Phil Venables, CISO of Google Cloud. stated in the publication.

The effort builds on Google’s experience developing cybersecurity models, including the collaborative Supply-chain Levels for Software Artifacts (SLSA) framework and BeyondCorp, a zero-trust architecture used by many organizations. .

Specifically, SAIF was designed to mitigate the risks inherent in AI systems, such as model theft, data poisoning of training data, malicious input via prompt injection, and extraction of sensitive information within training data. It’s the “first step”.

SAIF is built on six core principles:

1. Extending a Strong Security Foundation to the AI ​​Ecosystemincluding leveraging secure-by-default infrastructure protections (such as SQL injection mitigation techniques)
2. Expand detection and response to bring AI to your organization’s threat space: Monitor the inputs and outputs of generative AI systems to detect anomalies and use threat intelligence to predict attacks
3. Automate defenses to meet existing and emerging threats
4. Harmonize platform-level controls to ensure consistent security across your organizationstarting with the Google-owned Vertex AI and Security AI Workbench, the Perspective API is a free, open-source API developed by Google’s Jigsaw and Counter Abuse Technology team that uses machine learning to identify online “harmful” Identifies the comment.
5. Adapt controls to tune mitigations and create faster feedback loops for AI adoptionThis includes techniques such as reinforcement learning based on incidents and user feedback, updating training datasets, and fine-tuning models to strategically respond to attacks and red team exercises.
6. Understand the AI ​​system risk landscape in the surrounding business process By conducting an end-to-end risk assessment related to how an organization deploys AI

“We will soon publish several open-source tools to help implement the SAIF elements for AI security,” said Hansen and Venables.

They also vowed to expand Google’s BugHunter program to reward and encourage research into AI safety and security.

Read more: Ethical Hackers Could Earn Up to $20,000 for Finding ChatGPT Vulnerability

Finally, Google will help develop international standards for AI security, such as the National Institute of Standards and Technology (NIST) AI Risk Management Framework, Cybersecurity Framework, ISO/IEC 42001 AI Management System and ISO. said he was trying. /IEC 27001 Security Management Systems Standard.