Universal　steganalysis　include　feature　extraction　and　steganalyzer　design.　Most　universal　steganalysis　use　Support　Vector　Machine　(SVM)　as　steganalyzer.　However,　most　SVM-based　universal　steganalysis　are　not　to　be　very　much　effective　at　lower　embedding　rates.　The　reason　why　selective　SVMs　ensemble　improve　the　generalization　ability　was　analyzed,　and　an　algorithm　to　select　a　part　of　individual　SVMs　according　to　their　difference　to　build　the　ensemble　classifier　was　proposed,　which　based　on　the　selected　ensemble　theory-Many　could　be　better　than　all.　In　this　paper,　the　selective　SVMs　ensemble　algorithm　was　used　to　construct　a　strong　steganalyzer　to　improve　the　performance　of　steganographic　detection.　The　twenty　five　experiments　on　the　benchmark　with　2000　different　types　of　images　show　that:　for　popular　steganography　methods,　and　under　different　conditions　of　embedding　rate,　the　average　detection　rate　of　proposed　steganalysis　method　outperforms　the　maximum　average　detection　rate　for　the　steganalysis　method　based　on　single　SVM　with　improving　by　3.05%-12.05%;　and　for　the　steganalysis　method　based　on　BaggingSVM　with　improving　by　0.2%-1.3%.　©　(2012)　Trans　Tech　Publications,　Switzerland.