Abstract— In this paper, we present a novel universal approach which consists in exploring statistics in the compressed frequency domain. This approach is motivated by two main characteristics of the lossless compression step of the JPEG format. First, this step can be considered as a bijective mapping and then, when only few bits are flipped at its input, half the bits are flipped at the output. These properties, combined with a binary entropy deviation we pointed out, enable the design of detection schemes which the efficiencies are constant and do not depend in practice on the amount of information that has been embedded. These characteristics define a new class of promising functions for steganalysis. We illustrate our technique by considering RLE plus Huffman as such a function and design a new efficient universal steganalytic scheme to blindly detect the use of Outguess, F5 and JPHide and JPseek. Experimental results show that our steganalysis scheme is able to efficientl...