Making Experiments Dependable

摘要

In computer science and computer security we often do experiments to establish or compare the performance of one approach vs another to some problem, such as intrusion detection or biometric authentication. An experiment is a test or an assay for determining the characteristics of the item under study, and hence experimentation involves measurements.Measurements are susceptible to various kinds of error, any one of which could make an experimental outcome invalid and untrustworthy or undepend-able. This paper focuses on one kind of methodological error - confounding -that can render experimental outcomes inconclusive, but often without the investigator knowing it. Hence, valuable time and other resources can be expended for naught. We show examples from the domain of keystroke biometrics, explaining several different examples of methodological error, their consequences, and how to avoid them.