Several key exchange methods for wireless channels have been proposed in the literature. They are referred to as PHY-layer security techniques and are usually based on the channel's fading characteristics and the principle of reciprocity. In this paper, we present a novel PHY-layer security algorithm whose function is based on neural networks. Specifically, we present a full key exchange scheme which includes channel sampling and thresholding and neural network based error reconciliation. The proposed method's performance and offered security are studied through simulations and interesting conclusions are drawn about its overall utility.