In view of the problems in the existing feature fusion based JPEG steg analysis schemes,such as high redundancy in selected features and weak universality,a universal JPEG steganalysis approach based on improved boosting feature selection (BFS) method is presented. Feature redundancy is reduced in aspects of linear and nonlinear correlations.Statistical performance including auto-correlation coefficients and mutual information is introduced in feature evaluation rules. The algorithm of computing feature weighting is redesigned.The feature evaluation function of BFS is impro ved.Three complementary sets of features that have high detection accuracy are fused using the improved BFS algorithm.The selected optimal feature subset is used for training classifiers. Experiments are done in various embedding rates for three steganographic schemes with high concealment,i ncluding F5,Outguess and MME3.The results show that the detection accuracy of the proposed scheme is higher than that of some existing JPEG steganalysis approaches and some classical fusion methods. The fused features by improved BFS have lower correlation and this scheme has gr eater universality.