Nowadays,　there　exist　a　number　of　challenges　in　high-dimensional　features　space　for　steganalysis,　which　mainly　focus　on　the　difficulty　to　train　classifiers　and　high　computational　complexity.　Therefore,　we　propose　the　improved　RSM　algorithm　based　on　ensemble　pruning　for　high　dimensional　steganalysis.　Firstly,　SFS　algorithm　is　adopted　to　select　the　features　with　high　classification　ability　as　fixed　features,　while　the　remaining　features　are　selected　randomly　in　the　other　feature　space.　Secondly,　the　feature　subset　is　established　using　fixed　features　and　the　randomly　selected　features.　It　is　then　used　to　train　the　base　classifiers　afterwards.　Finally,　the　pruned　ensemble　is　obtained　to　yield　the　final　decision　using　FP-Tree　algorithm,　in　which　a　FP-Tree　is　built　to　compact　the　prediction　results　of　all　base　classifiers　for　the　validation　set　and　the　ensemble　with　the　best　predictive　accuracy　for　the　validation　set　is　output.　Experiment　results　demonstrate　that　the　proposed　algorithm　gains　the　small-size　pruned　ensemble　and　achieves　better　detection　performance　with　a　relatively　lower　computation　overhead　when　compared　with　the　traditional　ensemble　pruning　algorithms　such　as　Forward　Selection　and　Oriented　Order　against　HUGO　steganography.