Federated　Learning　(FL)　is　a　distributed　paradigm　enabling　clients　to　train　a　global　model　collaboratively　while　protecting　client　privacy.　During　the　FL　training　process,　the　statistical　heterogeneity　between　different　clients　can　compromise　the　overall　performance　of　the　global　model　and　its　generalization　ability　on　each　client,　making　it　difficult　for　the　training　process　to　converge.　This　paper　proposes　an　efficient　clustered　FL　(cFL)　method　called　FedCC,　which　aims　to　cluster　clients　based　on　their　inference　results　on　a　public　dataset.　As　inference　results　may　leak　client　data　distribution,　we　use　Locality　Sensitive　Hashing　(LSH)　to　transform　inference　results　into　Inference　Hash　Codes　(IHC),　which　are　irreversible　but　can　be　used　for　similarity　calculations.　The　server　compares　the　similarity　of　IHCs　between　clients　and　implements　dynamic　clustering　using　the　DBSCAN　algorithm.　FedCC　also　provides　an　elegant　method　to　quickly　select　the　appropriate　cluster　model　for　clients　without　downloading　all　cluster　models.　We　evaluated　FedCC　on　four　commonly　used　datasets　and　compared　them　against　seven　baselines.　Experimental　results　show　that　FedCC　achieves　faster　convergence　than　other　baselines　while　achieving　an　accuracy　1.66%　higher　than　the　state-of-the-art　baseline.　Finally,　we　further　validated　the　robustness　of　FedCC　against　Byzantine　attacks,　where　malicious　clients　upload　negative　gradients　to　reduce　model　accuracy　and　prevent　convergence.　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.　2024.