Device-free identification (DFI) is a promising technique, which could recognize human identity using his/her unique influence on surrounding wireless signals in a device-free and contact-free manner. It could maintain the privacy of a user and enable smart applications to provide customized service for a specific user. Despite its advantages over other person's identification systems, one fundamental problem to solve is that the accuracy of the DFI system is a little bit low due to the extremely noisy wireless measurements. The goal of this work is to explore and exploit a method to extract intrinsic features from the noisy channel state information (CSI) so as to realize high-performance DFI. To this end, we propose a novel empirical-mode-decomposition-based general DFI framework, which decomposes raw noisy CSI measurements into intrinsic mode functions (IMF) and extracts intrinsic features from the IMF components accordingly. Under the proposed framework, we also develop two DFI systems based on the respiration and gait biometric features. Extensive experiments carried out on commodity WiFi reveal that the developed systems could identify a person with an accuracy of over 90% from a group of ten persons.