This　paper　presents　a　new　model　for　intrusion　detection　based　on　hierarchical　neural　networks.　The　model　uses　techniques　of　protocol　analysis　and　data　mining　to　construct　features　automatically.　A　new　method　with　meta-learning　is　discussed,　which　divides　the　intricate　task　of　intrusion　detection　into　several　single　ones　and　equips　intrusion　detection　system　with　better　capacity　of　protecting　from　attack.