Dangerous　driving　behavior　detection　is　one　of　the　most　important　researches　in　Intelligent　Transportation　System　(ITS),　which　can　effectively　reduce　the　probability　and　number　of　traffic　accidents.　Although　some　recent　approaches　combined　with　deep　learning　techniques　have　been　proposed　for　detecting　dangerous　driving　behaviors,　the　protection　of　user＇s　privacy　is　neglected.　Therefore,　we　propose　a　Federated　Deep　Attention　Fusion　model　(FedDAF)　to　address　the　dual　security　issues　in　dangerous　driving　behavior　detection,　i.e.,　data　security　and　traffic　security.　On　the　Client　side,　we　design　the　Deep　Attention　Fusion　Network　for　extracting　and　learning　driving　process　features　as　well　as　fusing　the　environmental　factors　of　the　vehicle　in　driving.　On　the　Server　side,　the　Singular　Spectrum　Entropy　Aggregation　method　is　designed　to　aggregate　Clients　with　high　relevance　and　multiple　information　content,　thereby　realizing　safety　information　sharing　among　Clients.　Finally,　the　experimental　results　on　real　datasets　show　that　the　FedDAF　method　has　the　best　performance　on　several　evaluation　metrics　relative　to　the　existing　two　categories　of　benchmark　methods.