This paper presents a kernel-based LMS（KLMS） multi-user detection method. Kernel-based LMS detection projects the received signal of DS-CDMA system to high-dimensional character space by a Gaussian kernel function and then uses linear detection. By using the kernel trick, all calculation is in the low-dimensional space so that the complex calculation in high-dimensional is avoided. It is a nonlinear multi-user detector for original space data and the performance is closer to the performance of the optimum detector. Simulation results show that KLMS provides fast convergence and the same stable error compared with other LMS of variable step size by choosing a suitable kernel parameter.