This paper presents a new blind spectrum sensing (SS) algorithm based on a machine learning model: the radial basis function support-vector machines (RBF-SVM). As features, the introduced approach uses statistical tests that are based on the eigenvalues of the received signals covariance matrix. Since the decision on the frequency resource occupancy is in fact an issue of labeling binary data, SVM is intended as a potential technique for SS paradigm. The flexibility of SVM for linearly non-separable and high dimensional data makes it a good candidate for our issue, particularly that we consider low signal to noise ratios (SNR). Computer simulations shows that the proposal outperforms classical non-cooperative SS algorithms.