Jointing　multi-source　data　for　model　training　can　improve　the　accuracy　of　neural　network.　To　solve　the　raising　privacy　concerns　caused　by　data　sharing,　data　are　generally　encrypted　and　outsourced　to　a　group　of　cloud　servers　for　computing　and　processing.　In　this　client-cloud　architecture,　we　propose　FPPNet,　a　fast　and　privacy-preserving　neural　network　for　secure　inference　on　sensitive　data.　FPPNet　is　deployed　in　three　cloud　servers,　who　collaboratively　execute　privacy　computing　via　three-party　arithmetic　secret　sharing.　We　develop　the　secure　conversion　method　between　additive　shares　and　multiplicative　shares,　and　propose　three　secure　protocols　to　calculate　non-linear　functions,　such　as　comparison,　exponent　and　division　that　are　superior　to　prior　three-party　works.　Some　secure　modules　for　running　convolutional,　ReLU,　max-pooling　and　Sigmoid　layers　are　designed　to　implement　FPPNet.　We　theoretically　analyze　the　security　and　complexity　of　the　proposed　protocols.　With　MNIST　dataset　and　two　types　of　neural　networks,　experimental　results　validate　that　our　FPPNet　is　faster　than　the　related　works,　and　the　accuracy　is　the　same　as　that　of　plaintext　neural　network.　©　2022,　ICST　Institute　for　Computer　Sciences,　Social　Informatics　and　Telecommunications　Engineering.