People＇s　increasing　demand　for　high-quality　network　services　has　prompted　the　continuous　attention　and　development　of　network　traffic　classification　(NTC).　In　recent　years,　deep　flow　inspection　(DFI)　is　considered　to　be　the　most　effective　and　promising　method　to　solve　the　NTC.　However,　DFI　still　cannot　effectively　address　the　problem　of　changes　in　flow　characteristics　of　complex　packet　flows　and　the　discovery　of　new　traffic　categories.　In　this　paper,　we　propose　a　metric　learning　based　deep　learning　solution　with　feature　compressor,　named　deep　flow　embedding　(DFE).　The　feature　compressor　is　used　to　compress　the　feature　information　transmitted　layer　by　layer　in　DL　backbone　while　maintaining　the　computational　accuracy,　so　that　the　backbone　can　remove　as　much　noise,　redundancy,　and　other　irrelevant　information　from　the　input　data　as　possible,　and　achieve　more　robust　feature　extraction　of　network　traffic　flow.　The　deep　learning　(DL)　backbone　generates　an　embedding　vector　for　each　network　packet　flow.　Then　the　embedding　vector　is　compared　with　the　vector　template　preset　for　each　traffic　type　in　the　template　library　to　determine　the　category　of　the　packet　flow.　Experimental　results　verify　that　our　method　is　more　effective　than　the　traditional　DFI　methods　in　overcoming　the　problems　of　flow　characteristics　variation　and　new　category　discovery.　　©　2013　IEEE.