The objective of quantitative steganalysis is to achievereliable estimation of embedded message length of a suspected digitalobject. This type of methods has received considerable attention due toits capability of providing more detailed information about embeddedsecrets rather than just determining whether a suspected object isstego or not. In this paper, we present the methodology of "universalquantitative steganalysis", which is a practical, unified approach fordesigning quantitative steganalytic methods based on statisticallearning techniques. This methodology models the relation betweenembedded message length and statistical feature change caused by theembedding process with a multivariable function, and solves the problemof optimal parameter estimation with SVR (Support vector regression)technique. Experimental results indicate that new quantitativesteganalytic methods applying the presented methodology can achieveexcellent performance for F5 and MB1 steganographic mechanisms.