With　the　continuous　development　of　mobile　devices,　mobile　applications　bring　a　lot　of　convenience　to　people＇s　lives.　The　abuse　of　mobile　device　permissions　is　prone　to　the　risk　of　privacy　leakage.　The　existing　detection　technology　can　detect　the　inconsistency　between　the　declared　authority　and　the　actual　use　authority.　But　using　the　third-party　privacy　policy　as　the　analysis　basis　for　SDK　permissions　will　result　in　a　large　set　of　extracted　declaration　permissions,　which　will　lead　to　identifying　risky　applications　as　normal　applications　during　consistency　comparison.　The　prevailing　approach　involves　utilizing　models　based　on　TextCNN　to　extract　information　from　privacy　policies.　However,　the　training　of　TextCNN　relies　on　large-scale　annotated　datasets,　leading　to　high　costs.　This　paper　uses　BERT　as　the　word　vector　extraction　model　to　obtain　private　phrases　from　the　privacy　policy.　And　then　we　use　cosine　similarity　to　automatically　filter　permission　phrase　samples,　reducing　the　workload　of　manual　labeling.　On　the　other　hand,　existing　methods　do　not　support　the　analysis　of　Chinese　privacy　policies.　In　order　to　solve　the　problem　of　consistency　judgment　between　Chinese　privacy　policy　and　permission　usage,　we　implement　a　BERT-based　Android　privacy　policy　and　permission　usage　consistency　analysis　engine.　The　engine　first　uses　static　analysis　to　obtain　the　permission　list　of　Android　applications,　and　then　combines　the　BERT　model　to　achieve　consistency　analysis.　After　functional　and　speed　testing,　we　found　that　the　engine　can　successfully　run　the　consistency　analysis　function　of　Chinese　declaration　permissions　and　usage　permissions,　and　it　is　better　than　the　existing　detection　methods.