Machine learning has been used to detect a lot of malicious behavior, such as online auction fraud, fake reviews, spam, credit card fraud, and more. Therefore, researchers are trying to find solutions for making machine learning work when people are actively working to defeat it. Dr. Daniel Lowd, of the University of Oregon, uses game theory and machine learning to develop new algorithms that learn more robust classification models. By modeling the problem as a game between the machine learning system and the adversaries working to defeat it, Dr. Lowd and his team can use game theory to find an optimal strategy for each player, giving them an optimally robust classifier. With the help of game theory, Dr. Lowd hopes to make machine learning robust enough to apply to adversarial domains thereby making the Internet safer.

Dr. Lowd is particularly interested in learning robust classifiers for complex problems with many related predictions, such as labeling spammers in a social network or fake reviews in a network of user and product reviews. He hopes to develop a better understanding of the vulnerabilities of our current models, the weaknesses that attackers can exploit, and how to fix them. Dr. Lowd is collaborating with industry to apply and evaluate his methods in real-world settings. Finally, Dr. Lowd hopes to learn about the adversaries affecting computer learning as well as by seeing how spammers change their emails over time, researchers may be able to learn what spammers' weaknesses are and build models that exploit those as well. Machine learning has already had a huge impact and it has tremendous promise for the future. But in order to be effective, machine learning must be more robust; Dr. Lowd is working to make that happen.

Current research includes:

• Untrickable Predictive Models: Dr. Lowd is creating algorithms for learning predictive models that are hard to evade. By thinking like the attackers, he builds models that perform well against their future behaviors, not just their previous actions.

• Understanding Adversaries: Dr. Lowd hopes to understand the attackers' goals and costs. Better attacker models are important for building more robust predictive models. Dr. Lowd builds these models by observing adversaries' behaviors and patterns in data.

• Theoretical Connections: Dr. Lowd is looking at the theoretical connection between robustness to adversaries and robustness to other kinds of changes in data. This connection could lead to better predictive models in both adversarial and non-adversarial settings.