Internet-of-Things　(IoT)　devices　are　increasingly　used　in　people＇s　lives　and　production　in　various　industries.　To　detect　and　defend　against　Denial-of-Service　(DDoS)　attacks　that　occur　on　IoT　networks,　a　lot　of　methods　based　on　machine　learning　and　deep　learning　have　been　proposed　in　recent　years.　However,　these　methods　usually　do　not　consider　the　limitation　of　computational　resources　of　IoT　devices.　In　this　paper,　we　propose　an　edge　model　DDoS-Attack-Guard　(DAG)　based　on　Bi-GRU　and　ShuffleNet　for　DDoS　identification　and　classification　with　the　target　of　lightweight　and　real-time.　To　demonstrate　the　performance　of　our　models,　we　use　the　CICDDoS2019　dataset　to　test　the　identification　and　classification　accuracy　as　well　as　the　model　inference　time.　In　addition,　we　build　a　multi-layer　coder-decoder　structure　that　can　extract　the　potential　temporal　contextual　features　of　DDoS　traffic,　and　introduce　a　reconstruction　structure　that　can　improve　model　training.　Through　ablation　experiments　and　comparative　experiments,　our　model　has　an　average　inference　speed　of　2.5ms　across　different　data　sizes,　which　is　50%　faster　than　the　Sota　method,　while　hitting　99.3%　and　99.9%　accuracy　in　identification　and　classification　respectively.