DEF CON 26 AI VILLAGE - Ariel Herbert Voss - Machine Learning Model Hardening For Fun and Profit

Machine learning has been widely and enthusiastically applied to a variety of problems to great success and is increasingly used to develop systems that handle sensitive data - despite having seen that for out-of-the-box applications, determined adversaries can extract the training data set and other sensitive information. Suggested techniques for improving the privacy and security of these systems include differential privacy, homomorphic encryption, and secure multi-party computation. In this talk, we’ll take a look at the modern machine learning pipeline and identify the threat models that are solved using these techniques. We’ll evaluate the possible costs to accuracy and time complexity and present practical application tips for model hardening. I will also present some red team tools I developed to easily check black box machine learning APIs for vulnerabilities to a variety of mathematical exploits.

[karlfimm]

Excellent information, badly presented.

[aazz]

Wait, doesn't she mean non-convex optimization (I wish it were convex). Am I missing something?