This thesis presents the first comprehensive study and new methods for radiometric fingerprinting of the Cognitive Radio (CR) devices. The scope of the currently available radio identification techniques is limited to a single radio adjustment. Yet, the variable nature of the CR with multiple levels of parameters and adjustments renders the radiometric fingerprinting much more complex. We introduce a new method for radiometric fingerprinting that detects the unique variations in the hardware of the reconfigurable radio by passively monitoring the radio packets. Several individual identifiers are used for extracting the unique physical characteristics of the radio, including the frequency offset, modulated phase offset, in-phase/quadrature-phase offset from the origin, and magnitude. Our method provides stable and robust identification by developing individual identifiers (classifiers) that may each be weak (i.e., incurring a high prediction error) but their committee can provide a strong classification technique. Weighted voting method is used for combining the classifiers. Our hardware implementation and experimental evaluations over multiple radios demonstrate that our weighted voting approach can identify the radios with an average of 97.7% detection probability and an average of 2.3% probability of false alarm after testing only 5 frames. The probability of detection and probability of false alarms both rapidly improve by increasing the number of test frames.