Current　SVM-based　image　steganographic　detection　algorithmins　haven＇t　considered　the　impact　of　specific　data,　and　the　choice　of　the　parameter　greatly　affects　the　classification　performance,　it＇s　necessary　to　constructing　the　kernel　function　from　the　perspective　of　specific　data.　This　paper　proposes　a　steganographic　detection　method　for　JPEG　image　that　base　on　the　data-dependent　concept,first　obtain　the　initial　classifier　by　SVM　training,　then　the　kernel　function　is　modified　with　conformal　transformation　by　using　the　information　of　Support　Vectors,　re-train　with　the　new　kernel　to　enlarge　the　spacing　around　classfication　boundary,　iterate　until　getting　the　best　result.　Experimental　results　illustrate　our　method　dose　effectively　improve　the　classification　accuracy　of　image　universal　steganalysis,　futhermore,　a　high　classification　accuracy　under　the　default　parameters　makes　the　algorithmin　more　practical.　©　2010　IEEE.