Abstract: The Antibody Network (ABNET), which is a new Artificial Neural Networks (ANN) based on immune principle, has been proved to have good ability of unsupervised and competitive learning in experiments. But there is no precedent to use ABNET in practical engineerings because the present researches about it are still focusing on principle and experiments. While the good qualities of ABNET are reserved, such as structure generated automatically, unsupervised and competitive learning based on clone selection and mutation,the way of initiating ABNET, expressing a antibody, and updating structure is improved properly. The mail filter based on improved ABNET is better at adaptation and accuracy in experiments, compared with traditional mail filters.