In　recent　years,　the　release　of　data　value　under　privacy-preserving　has　always　been　the　research　focus.　As　an　essential　data　protection　technology,　encrypted　computing　provides　a　good　solution.　Based　on　this,　a　federated　learning　model　is　first　constructed　to　meet　the　needs　of　participants　using　local　data　for　joint　modeling.　Secondly,　the　meta-learning　method　accelerates　the　federated　learning　model＇s　training　effect　and　improves　the　accuracy　of　model　detection.　Finally,　the　variational　auto-encoder　is　used　to　optimize　the　federated　learning　model,　which　speeds　up　the　training　of　the　federated　learning　model　and　improves　the　system＇s　overall　security.　Experimental　results　show　that　the　proposed　system　can　provide　secure　computing　conditions　for　each　data　provider　and　better　performance.　©　2022　Institute　of　Physics　Publishing.　All　rights　reserved.