The　most　of　Intrusion　detection　systems　divided　data　into　two　classes,　which　were　normal　and　abnormal,　so　that　it　might　lose　some　important　information.　The　goal　of　feature　selection　was　to　decrease　the　redundant　features　for　anomaly　detection,　and　maintain　the　same　high　accuracy　as　the　original　features.　It　proposed　an　anomaly　intrusion　detection　technique　based　on　feature　selection　and　multi-class　support　vector　machines(SVM).　The　feature　selection　method　merged　RS,　SVDF,　LGP　and　MARS.　Then,　data　was　divided　into　five　classes　by　the　multi-class　SVM.　The　experimental　results　demonstrate　that　the　false　positive　rate　of　DoS　is　the　highest　one　among　four　methods.