The　research　on　community　detection　is　usually　based　on　the　topological　structure　and　attribute　information　of　complex　networks　to　improve　computation　precision.　However,　as　more　and　more　people　pay　attention　to　the　disclosure　of　personal　privacy,　detecting　communities　without　leaking　sensitive　information　has　become　a　hot　topic　in　complex　network　analysis.　In　this　paper,　we　first　propose　a　distributed　privacy-preserving　graph　learning　model.　Second,　we　develop　a　multi-label　propagation　algorithm　(MLPA)　based　on　the　model　to　detect　overlapping　communities　securely　on　the　horizontally　distributed　networks　with　attributes.　A　novel　perturbation　strategy　is　combined　with　homomorphic　encryption　to　achieve　flexible　privacy　control　and　strict　privacy　protection.　Moreover,　a　node　similarity　calculation　method　is　proposed　to　consider　the　structural　and　attribute　influences　of　each　node＇s　neighbors　in　label　propagation　no　matter　the　attributes　are　numeric　or　categorical.　The　experiments　on　real-world　and　artificial　networks　demonstrate　that　our　algorithm　achieves　identical　results　as　the　standaloneMLPAand　higher　accuracy　(200%)　than　the　simple　distributed　MLPA　without　federated　learning.