Adversarial Classification When Even Good Types Can Fake
6 Pages Posted: 14 Jan 2009
Date Written: December 8, 2007
Abstract
In some classification domains, firms face agents who actively manipulate their information to mislead the firm about their true types so as to avoid unfavorable decisions as a result of the classification. In such domains, firms should take the possibility of applicants' faking behavior into consideration in their decision making. We consider situations where the firm faces agents who can modify instances regardless of their type; unlike prior work, we don't restrict ourselves to those situations where only malicious agents manipulate their data. We show that the firm is never better off when agents have the ability to fake than when they do not. However, surprisingly, a reduction in faking cost does not always hurt the firm, implying that a firm may sometimes prefer an environment in which agents can fake more easily over another in which it is more difficult to fake.
Suggested Citation: Suggested Citation