The Dark Side of AI: The Hidden Supply Chain Risks in Open... - Jossef Kadouri & Tzachi Zornshtain
Published 2024-06-27
There's a rise in powerful AI tools. Are they secure enough? In this talk, I will demonstrate how open source AI models are built, with a focus on their weak spots. I will then give a live demo of how to exploit these weaknesses to build malicious models and show how easy it is to publish them to HuggingFace. I will show the following techniques: - embedding malicious code in models - data poisoning for existing models and training to target specific victims