We proposed a blind separation method based on second-order statistics for hopping-frequency signals. Firstly, a cost function was constructed to estimate separation matrix by exploiting the properties of non-stationary and non-whitening of signals, then updated the separation matrix by gradient steepest method, so as to obtain optimal separation matrix. The simulation results show that the method can separate multiple-hopping frequency signals off- or on-line, and its superior robustness performance to the conventional methods.