Adversarial Classification: Impact of Agents’ Faking Cost on Firms and Agents

Classification of agents such as suppliers and customers into good and bad types based on their attributes is ubiquitous in business. In adversarial classification contexts, firms face agents who fake their attributes to receive favorable decisions from the classifier; for instance, suppliers could misrepresent their safety or quality features in order to qualify as a firm’s accredited vendor. Anticipating such faking by agents, firms may strategically design their classifiers and may also verify the information presented by agents. We examine the impact of agents’ faking on the firm and on agents in a context in which agents of two types—good‐risk and bad‐risk—have possibly different faking costs. We find that the firm is hurt when bad‐risk agents’ faking costs decrease, and the firm benefits when good‐risk agents’ faking costs decrease. We find that agents do not always benefit when faking becomes easier. Further, a decrease in faking costs of one type will sometimes benefit the other type. We show that when the faking costs decrease the firm may be better off using some attributes that do not offer a positive value to the firm at the higher faking cost. Though supplementing a strategic classifier with verification of the agents’ information could benefit the firm, 100% verification may not be optimal even if there was no cost incurred in conducting verification. Experiments conducted using a Naive Bayes classifier on a real‐world dataset demonstrate that our findings hold for extant (and potentially non‐optimal) commercial classifiers.