The aim of this paper is to present a method to improve the detection accuracy of a blind steganalysis system from an engineering perspective, enlightening the effective methods to make a good dataset to be usable in building a blind steganographic system and showing significant results related to the use of the wavelet coefficients? statistics. Our experiments have been made using a large dataset of JPEG images obtained from a certified source that can be offered to the scientific community to compare the different proposed approaches in an objective way. The results show the effectiveness of the used solution on a large dataset and the real improvements in the detection accuracy.