search menu icon-carat-right cmu-wordmark

Poster - Train, but Verify: Towards Practical AI Robustness

November 2020 Poster
Nathan M. VanHoudnos, Jon Helland

This presentation describes efforts to train AI systems to enforce at least two security policies and verify security by testing against realistic threat models.

Publisher:

Software Engineering Institute

Abstract

In this “Train, but Verify” project, we attempt to address the gap in the state of the art on secure training of machine learning (ML) systems with two objectives:

  1. Train secure artificial intelligence (AI) systems by training ML models to enforce at least two security policies.
  2. Verify the security of AI systems by testing against declarative, realistic threat models.