Universal　steganalysis　techniques　attempt　to　detect　hidden　information　without　knowledge　about　the　steganographic　methods.　One　of　the　most　important　things　is　to　find　feature　sets,　which　are　sensitive　to　the　embedding　process.　Whether　these　features　are　＇good＇　will　directly　influence　the　accuracy　of　detection.　This　paper　describes　an　approach　to　define　sensitive　feature　sets　using　ICA　(independent　component　analysis)　decomposition　and　prediction　in　order　to　build　statistical　models　of　image　independent　component.　Kernel-SVM　is　then　used　to　discriminate　between　stego-images　and　cover-images.