Adversarial Classification: Data Mining, Statistical Learning, and Data Visualization for Complex Data

Research Areas: Assured Identity and Privacy,

Principal Investigator: Bowei Xi, Bill Cleveland

Adversarial classification applications range from the annoyance of spam to the damage of computer hackers to the destruction of terrorists. In all of these cases, statistical classification techniques play an important role in distinguishing the legitimate from the destructive. These problems pose a significant new challenge not addressed in previous research: The behavior of a class controlled by the adversary may adapt to avoid detection. Hence the future datasets and the training data are no longer from the same populations. We model the problem as a two-player game, where the adversary tries to maximize its return and the data miner tries to minimize the misclassification error. We examine under which conditions an equilibrium would exist, and provide a method to estimate the classifier performance and the adversary’s behavior at such an equilibrium point — the players’ equilibrium strategies. Such information is critical for the construction of a resilient classifier against the adversary.

Personnel

  • Jongwoo Song

Keywords: data mining, statistics, classification, visualization