In this paper, we study channel-based authentication, where the receiver can identify and authenticate the senders through channel vectors estimated from their frames. The authentication process is formulated as a sequence of hypothesis test problems. In order to improve the detection probability and reduce the false alarm probability, two schemes are proposed based on different classification algorithms in machine learning. Specifically, support vector machine (SVM) based authentication schemes and the linear Fisher discriminant analysis (LFDA) based authentication scheme are proposed by exploiting three channel features, including the time-of-arrivals, received signal strengths, and cyclic-features of the channels. In SVM based schemes, the linear and nonlinear SVMs are used to generate classifiers to solve the hypothesis test problems. In LFDA based scheme, a linear combination of these three channel features is used as the test statistic, which is compared with a threshold to perform authentication. Simulation results demonstrate that the proposed schemes perform better in terms of the misdetection probability and the false alarm probability than several existing typical channel-based authentication schemes. Moreover, the time complexity and space complexity of the proposed schemes are analyzed, and the LFDA based scheme performs the best.