With　the　dramatic　opening-up　of　network,　network　security　becomes　a　severe　social　problem　with　the　rapid　development　of　network　technology.　Intrusion　Detection　System　(IDS)　is　an　innovative　and　proactive　network　security　technology,　which　becomes　a　hot　topic　in　both　industry　and　academia　in　recent　years.　There　are　four　main　characteristics　of　intrusion　data　that　affect　the　performance　of　IDS　including　multicomponent,　data　imbalance,　time-varying　and　unknown　attacks.　We　propose　a　novel　IDS　framework　called　HMLD　to　address　these　issues,　which　is　an　exquisite　designed　framework　based　on　Hybrid　Multi-Level　Data　Mining.　In　this　paper,　we　use　KDDCUP99　dataset　to　evaluate　the　performance　of　HMLD.　The　experimental　results　show　that　HMLD　can　reach　96.70%　accuracy　which　is　nearly　1%　higher　than　the　recent　proposed　optimal　algorithm　SVM+ELM+Modified　K-Means.　In　details,　HMLD　greatly　increased　the　detection　accuracy　of　DoS　attacks　and　R2L　attacks.