In　this　paper,　we　explore　the　issues　of　wireless　communication　physical　layer　security　in　time　diversion　duplex　(TDD)　system　and　propose　a　directional　reactive　jamming　scheme　based　on　machine　learning.　The　proposed　scheme　can　recognize　the　identities　of　different　users　by　their　channel　state　information　(CSI)　using　the　methods　of　K-means　and　support　vector　machine　(SVM).　Besides,　more　suitable　features　of　CSI　have　been　chosen　for　K-means　clustering.　Further　more,　the　security　measures　are　taken　once　identifying　the　eavesdropper,　such　as　transmitting　interference　signal　from　BS　directly　to　the　eavesdropper.　After　theoretical　analysis,　we　derive　the　analytical　expressions　of　Secrecy　Capacity　and　Secrecy　Outage　Probability　(SOP)　based　on　the　scheme.　Finally,　Monte-Carlo　simulations　have　been　done　to　further　explore　the　performance　and　verity　our　analytic　results.　According　to　our　analytic　and　simulation　results,　the　recognition　method　of　our　scheme　performs　with　high　accuracy.　Meanwhile,　if　the　system　location　of　eavesdropper　is　not　coincide　with　the　legitimate　user,　the　system　can　always　obtain　high　security　level.　The　simulation　results　with　different　SINRs　also　show　that　our　scheme　is　robust　in　different　communication　environments.