Abstract: A convenient and secure management scheme of cipher key is proposed taking advantage of the associative memory function from neural networks and embedding the cipher key in the user's facial biometrics feature. Extracting the user's facial biometrics feature and discretizing it to be a binary sequence through the principal component analysis (PCA), and the key is integrated into the facial biometrics sequence via XOR (exclusive-OR), and the XOR result and some random sequences are used to train the generalized regression neural networks. Eventually, what a user has to do is just storing the trained parameters of neural networks, XOR result of the key and facial biometrics sequence without the key to be stored or memorized. Then, only if a user inputs his own face image the key can be restored from the stored data whenever the key is required. An attacker is impossible to get the key or the information on user's facial biometrics feature, even if he has obtained the stored data. The experiment results showed that the proposed scheme is feasible.