An encrypting algorithm based on RBF neural network is presented. The clear-text is regarded as RBF target vector. RBF network is trained with random input vector, and the trained results are regarded as the cryptograph. With the feature that the input vector of RBF neural network can be any random value, there are the infinite kinds of ways to build the input vector, so the cryptograph cannot be deciphered theoretically. The random seed used to generate the input vector is regarded as the secret-key. The receiver can get the clear-text from the cryptograph by reconstructing RBF neural network and its input vector with the secret-key. The simulation experiments prove the algorithm is feasible and effective.