Federated　learning　not　only　realizes　collaborative　training　of　models,　but　also　effectively　maintains　user　privacy.　However,　with　the　widespread　application　of　privacy-preserving　federated　learning,　poisoning　attacks　threaten　the　model　utility.　Existing　defense　schemes　suffer　from　a　series　of　problems,　including　low　accuracy,　low　robustness　and　reliance　on　strong　assumptions,　which　limit　the　practicability　of　federated　learning.　To　solve　these　problems,　we　propose　a　Robustness-enhanced　privacy-preserving　Federated　learning　with　scaled　dot-product　attention　(RFed)　under　dual-server　model.　Specifically,　we　design　a　highly　robust　defense　mechanism　that　uses　a　dual-server　model　instead　of　traditional　single-server　model　to　significantly　improve　model　accuracy　and　completely　eliminate　the　reliance　on　strong　assumptions.　Formal　security　analysis　proves　that　our　scheme　achieves　convergence　and　provides　privacy　protection,　and　extensive　experiments　demonstrate　that　our　scheme　reduces　high　computational　overhead　while　guaranteeing　privacy　preservation　and　model　accuracy,　and　ensures　that　the　failure　rate　of　poisoning　attacks　is　higher　than　96%.