Current　image　steganographic　detection　algorithms　are　unable　to　make　full　use　of　the　geometry　of　unlabeled　image　examples,　detection　performance　is　subject　to　a　few　labeled　examples,　which　is　utilized　for　training.　In　this　paper,　we　propose　an　effective　steganographic　detection　method　for　JPEG　image　that　rely　on　the　overall　dataset.　The　method　is　combined　with　semi-supervised　kernel　in　the　presence　of　unlabeled　examples.　Semi-supervised　kernel　method　constructs　data　adjacency　graph　to　obtain　Gram　matrix,　then　we　obtain　the　proposed　method　by　incorporating　graph　Laplacian　into　kernel-based　algorithms,　which　is　effective　integration　of　the　cluster　assumption　and　manifold　assumption.　Our　method　utilizes　the　geometry　of　all　examples　with　manifold　regularization　to　produce　smooth　decision　functions　and　thus　improving　the　performance　universal　steganographic　detection.　Experimental　results　show　the　effectiveness　of　our　proposed　method.　©　2010　IEEE.