Threat Modeling for AI/ML Systems

Introduction

• Threat modeling introduction
• What you should know

1. Threat Modeling Overview

• Threat modeling is important when building AI systems
• The four-question framework structures your work
• Anyone can threat model and you should, now
• Trustworthy AI: Threat modeling is better than principles

2. What Are You Working on with ML?

• ML for business, offense, defense, and software
• Draw your architecture
• Deployment architectures influence your threats
• Training data is a crucial variable
• The stochastic parrot

3. What Can Go Wrong with ML Security

• The OWASP Top Ten as a checklist
• The Berryville Institute Exhaustive List
• Microsoft's frameworks for security flaws
• Prompt injection
• Embarrassing and hostile results

4. What Can Go Wrong with AI: Trustworthiness

• NIST Framework
• EU's AI Act
• Current harms
• Scenarios

5. What Are You Going to Do about It?

• Specific frameworks
• Mitigations advance faster than threats
• Deploying new technology isn't a one-and-done

Conclusion

• Next steps