Tutorial: Applying Responsible AI Principles for Generative AI

This tutorial addresses the urgent need for responsible research and development in the rapidly evolving field of Generative AI (GenAI). Designed for AI researchers and practitioners, it offers a comprehensive exploration of Responsible AI (RAI) principles tailored specifically to Generative AI technologies. Participants will gain a deep understanding of the ethical implications of advancements in large language models and generative systems, including critical issues such as bias mitigation, privacy preservation and beyond. We go beyond theoretical discussions by offering practical, research-oriented strategies, innovative methodological frameworks and hands-on labs. Participants will learn techniques for detecting and mitigating biases in Generative AI models and gain experience applying Responsible AI principles in real-world research scenarios. 

• Need for Responsible AI in GenAI: Highlight the potential risks of GenAI amid its explosive growth and adoption.

• Responsible AI Principles: Explain the foundational principles of RAI

• Practical challenges faced in RAI: Highlight the vulnerabilities and challenges of adopting RAI in real-world use cases

• Mitigation approaches (ShieldGemma, LlamaGuard): Understand how these frameworks enhance the safety of LLMs

Lab 1: Protecting Sensitive Data in Gen AI Model Responses

In this lab, you will learn how to:

• Access a pre-created Jupyter notebook in a Vertex AI Workbench instance.

• Install Python packages for Vertex AI and Cloud Data Loss Prevention (DLP) API.

• Generate example text with sensitive data using the Gemini 2.0 Flash model.

• Define and run Python functions to redact different types of sensitive data in Gemini 2.0 Flash model responses using the DLP API.

Lab 2: Safeguarding with Vertex AI Gemini API

In this lab, you will learn how to:

• Call the Vertex AI Gemini API and inspect safety ratings of the responses.

• Understand the various types of security parameters.

• Define a threshold for filtering safety ratings according to your needs.

Lab 3: End-to-end solution with RAI guardrails

This lab highlights how to build an end-to-end solution while taking RAI guardrails into consideration, these include:

• Sending data to open models like Llama.

• Adding guardrails by implementing ShieldGemma or LlamaGuard.

• Evaluating content safety.

• Employ one class of safeguards—content classifiers for filtering.

Lab 4: [OPTIONAL] Differential Privacy in Machine Learning with TensorFlow Privacy

This lab helps you learn how to use differential privacy in machine learning using TensorFlow Privacy.

• Wrap existing optimizers into their differentially private counterparts using TensorFlow Privacy.

• Practice checking hyperparameters introduced by differentially private machine learning.

• Measure the privacy guarantee provided using analysis tools included in TensorFlow Privacy.