Toward Reliable JPEG Steganalysis (at QF100)

Contrary to classical steganalysis methods focused on the detection performance, this paper proposes a reliable steganalysis method targeting two goals: prescribing a given, potentially very small, False-Positive (FP) rate while maximizing the detection accuracy. This is the first step toward operational steganalysis where the forensics agent needs an accurate FP rate in order to make a decision. Because JPEG steganalysis at Quality Factor (QF) 100 is associated with an accurate and robust distribution of Cover images by computing the rounding error after JPEG decompression, we use this domain to derive statistical tests associated with theoretical FP rates. A Kolmogorov-Smirnov test applied on different pixel positions of the 8 × 8 block, associated with an aggregation strategy, and a block filtering pre-processing are combined to propose an accurate yet reliable detector for FP going below 10
⁻⁴
. Experiments are both assessed on synthetic images and the "wild" ALASKA database using J-UNIWARD.