Blind　steganalysis　detects　hidden　message　without　any　knowledge　about　steganographic　method.　To　implement　blind　steganalysis,　statistical　model　based　on　high-order　wavelet　decomposition　is　built　to　capture　statistical　difference　between　cover　images　and　stego　images.　However,　not　all　wavelet　statistics　are　able　to　reflect　well　statistical　changes　due　to　hidden　message　embedded.　Analysis　of　Variance　(ANOVA)　is　applied　to　test　which　wavelet　statistics　are　more　sensitive　to　hidden　message.　The　statistics　that　are　sensitive　to　hidden　message　are　selected　as　image＇s　features.　Using　Support　Vector　Machine　(SVM)　as　classifier,　our　experiment　results　show　that　ANOVA　is　effective　and　our　method　effectively　improves　accuracy.