Most　current　steganalysis　schemes　are　based　on　the　Cachin＇s　statistical　undetectability.　However,　the　lack　of　the　statistical　model　for　the　natural　images　the　performance　of　the　current　steganalyzers.　In　this　paper,　we　propose　a　novel　steganalyzer　for　JPEG　images　based　on　manifold　learning,　which　overcomes　the　statistical　steganalyzer＇s　deficiencies.　The　feature　extraction　in　this　steganalyzer　is　completed　by　the　nonlinear　dimensionality　reduction　method　(ISOMAP),　which　will　greatly　reduce　the　dimensionality　of　the　feature　space　without　influencing　the　performance　of　our　steganalyzer.　Experimental　results　show　the　effectiveness　of　this　method　and　also　demonstrate　the　promise　of　the　proposed　scheme　used　both　as　a　specific　and　a　blind　steganalyzer.