Aiming at the disadvantages of feature extraction and selection in the traditional method for passive millimeter-wave (MMW) metal target recognition, the existence and characteristics of low dimensional manifold of the short-time Fourier spectrum of metal target echo signal are explored using manifold learning algorithm, Laplacian eigenmaps. Target classification is performed through comparing the similarity of the test samples and the positive class in terms of the low dimensional manifold. The experiments show that the method gets higher recognition rate than other linear and kernel-based nonlinear dimensionality reduction algorithm, and is robust to data aliasing.