Verifiable Machine Learning for Security

In recent years, machine learning techniques have been increasingly applied to many critical problems in the cybersecurity domain, including detecting malware, spam, online fraud, hate speech, etc. However, there are many challenges to reliably deploy these solutions for security applications, since real-world adversaries are constantly trying to evade machine learning systems. My research focuses on solving this problem by increasing the cost for attackers to succeed.In this talk, I will discuss methods to train security classifiers with verified robustness properties. Robustness properties are security guarantees of the classifier that can eliminate certain classes of evasion attacks. I will show how to use security domain knowledge and economic cost measurement studies to formulate robustness properties to capture general classes of evasion strategies that are inexpensive for attackers. Then, I will describe new algorithms to train security classifiers to satisfy these properties. I will show how to apply the methods to detect PDF malware, Twitter spam, and Cryptojacking, and demonstrate that it is not only sound but also practical. My key result is, enforcing robustness properties can increase the economic cost of evasion. In the future, I want to integrate new machine learning models as a fundamental part to solve hard problems in security.