Live　video　hosted　by　streamers　is　being　sought　after　by　an　increasing　number　of　Internet　users.　Some　streamers　mix　pornographic　content　with　live　video　for　profit　and　popularity,　but　this　greatly　harms　the　network　environment.　To　effectively　identify　porn　streamers,　a　multilevel　fusion　method　of　multimodal　deep　features　for　porn　streamer　recognition　in　live　video　is　proposed　in　this　paper.　(1)　Visual　and　audio　features　including　spatial,　audio,　motion,　and　temporal　context　in　live　video　are　extracted　by　a　multimodal　deep　network.　(2)　Audio-visual　attention　features　are　obtained　by　fusing　visual　and　audio　features　at　the　feature　level　based　on　a　multimodal　attention　mechanism.　(3)　Text　features　are　extracted　by　using　the　bullet　screen　text　network　based　on　the　BERT　(bidirectional　encoder　representations　from　transformers)　model　after　collecting　text　information　from　the　viewers　bullet　screen　comments.　(4)　The　prediction　results　of　the　audio-visual　deep　network　and　the　bullet　screen　text　network　are　fused　at　the　decision　level　to　improve　the　porn　streamer　recognition　accuracy.　We　build　a　real-world　dataset　of　porn　streamers　and　conduct　experiments　and　demonstrate　that　our　method　can　improve　the　porn　streamer　recognition　accuracy.　©　2020　Elsevier　B.V.