The　defender　needs　to　predict,　detect　and　understand　attacks,　and　makes　good　decisions　about　defense　strategies.　Because　the　target　of　attackers　and　defenders　is　oppositional　and　their　strategies　are　interdependent,　the　selection　of　optimal　defense　strategy　is　a　complex　issue.　In　this　paper,　the　issue　of　optimal　defense　strategy　selection　is　defined　and　formalized.　A　new　attack-defense　stochastic　game　model　is　proposed　to　describe　the　offensive　and　defensive　conflict　of　attackers　and　defenders　in　network　security　and　address　the　issue　of　optimal　defense　strategy　selection.　The　model　is　a　dynamic　multi-player　and　multi-state　model　which　is　expanded　by　normal　attack-defense　game　and　Markov　decision　process.　By　viewing　privilege　state　in　node　of　attacker　as　elements　in　attack-defense　stochastic　game,　we　can　model　the　dynamic　transition　of　attack　and　defense　state　and　compute　the　probabilities　of　attacker　and　defender　behavior.　This　paper　analyzes　the　cost　factors　related　to　attack　and　defense　and　provides　a　cost-benefit　analysis　method　that　helps　defender　evaluate　and　select　defense　strategies.　An　algorithm　for　defense　strategy　selection　based　on　those　models　is　proposed.　A　representative　network　example　is　provided　to　illustrate　our　models　and　demonstrate　the　efficacy　of　our　models　in　the　prediction　of　attack　behaviors　and　decision　of　optimal　defense　strategies.