Gait authentication based on accelerometers is a nonintrusive biometric measurement. It is a novel and feasible way to enhance the security of portable electronic devices. To boost authentication performances, a decision-level data fusion algorithm via neural network is proposed in this paper. The proposed algorithm fuses acceleration signals in different directions and classical approaches for matching gait patterns such as correlation, Euclidean distance etc. In our experiments, data sets for training and test consist of 17, 20 subjects separately. The Equal error rate (EER) is cut to 0.82%, which is much lower than other proposed approaches so far.