Proactive　caching　is　a　novel　smart　communication　resource　management　method　that　can　offer　intelligent　and　economic　networking　services　in　the　next　generation　cellular　networks.　In　proactive　caching,　a　common　operation　is　using　k-means　to　estimate　content　popularity.　However,　during　the　process,　the　base　stations　have　to　collect　user＇s　location　and　content　preference　information　to　train　a　k-means　model,　which　causes　user　privacy　leakage.　And　current　privacy-preserving　k-means　schemes　usually　suffer　dramatic　user　quality　of　experience　reduction,　and　cannot　deal　with　the　user　dropout　condition.　Therefore,　we　propose　a　privacy-preserving　federated　k-means　scheme　(named　PFK-means)　for　proactive　caching　in　the　next　generation　cellular　networks.　PFK-means　is　based　on　two　privacy-preserving　techniques,　federated　learning　and　secret　sharing.　In　PFK-means,　a　suite　of　secret　sharing　protocols　are　designed　to　lightweight　and　efficient　federated　learning　of　k-means.　These　protocols　allow　privacy-preserving　k-means　training　for　proactive　caching　when　there　are　dropout　users.　We　seriously　analyze　the　security　of　PFK-means　and　conduct　comprehensive　experiments　to　prove　its　security,　effectiveness　and　efficiency.　Through　comparison,　we　can　conclude　that　PFK-means　outperforms　other　existing　related　schemes.　©　2020　Elsevier　Inc.