In modern cognitive ratio systems, the spectrum is becoming increasingly crowded and expensive; thus spectrum sensing becomes more important than ever before. Traditional spectrum sensing assumes Gaussian noise (or of other given distributions) in general. However when secondary users (SUs) have no prior information about the measurement distributions, the spectrum sensing schemes assuming given distribution forms (even if the parameters are assumed to be unknown) no longer apply. In this paper we propose a universal quickest change detection scheme based on density ratio estimation for spectrum sensing by detecting the sudden change of spectrum (e.g., the emergence of primary user), where neither the pre- change nor post-change distribution (even the distribution forms) is known to SUs, thus achieving robustness to complex spectrum environment.