EP84 How to Secure Artificial Intelligence (AI): Threats, Approaches, Lessons So Far

Guest:

• Alex Polyakov, CEO of Adversa.ai

Topics:

• You did research by analyzing 2000 papers on AI attacks released in the previous decade. What are the main insights?
• How do you approach discovering the relevant threat models for various AI systems and scenarios?
• Which threats are real today vs in a few years?
• What are the common attack vectors? What do you see in the field of supply chain attacks on AI, software supply, data?
• All these reported cyberphysical attacks on computer vision, how real are they, and what are the possible examples of exploitation? Are they a real danger to people?
• What are the main differences between protecting AI vs protecting traditional enterprise applications?
• Who should be responsible for Securing AI? What about for building trustworthy AI?
• Given that the machinery of AI is often opaque, how to go about discovering vulnerabilities? Is there responsible disclosure for AI vulnerabilities, such as in open-source models and in public APIs?
• What should companies do first, when embarking on an AI security program? Who should have such a program?

Resources:

• "EP52 Securing AI with DeepMind CISO" (ep52)
• "EP68 How We Attack AI? Learn More at Our RSA Panel!" (ep68)
• Adversarial AI attacks work on Humans (!)
• "Maverick* Research: Your Smart Machine Has Been Conned! Now What?" (2015)
• "The Road to Secure and Trusted AI" by Adversa AI
• "Towards Trusted AI Week 37 – What are the security principles of AI and ML?"
• Adversa AI blog
• AIAAIC Repository
• Machine Learning Security Evasion Competition at MLSec