Cost polarization by dequantizing for JPEG steganography

In this article, we study a recently proposed method for improving empirical security of steganography in JPEG images which the sender starts with an additive embedding scheme symmetrical costs ± 1 changes and then decreases cost one these based on image obtained by applying deblocking (JPEG dequantization) algorithm to cover JPEG. This approach provides rather significant gains at negligible complexity overhead wide range quality factors across various schemes. Challenging original explanation inventors idea, is interpreting dequantized as estimate precover (uncompressed) image, provide alternative arguments. The key observation main reason why works how polarizations individual DCT coefficients work together. By using MiPOD model content uncompressed show that polarization technique chances “bad” combinations would likely be introduced symmetric costs. statement quantified computing likelihood stego w.r.t. multivariate Gaussian distribution domain. Furthermore, it shown spatial discontinuities between blocks (blockiness) enforces desirable correlations blocks. To further prove point, source adheres model, simple Wiener filter can serve equally well deep-learning deblocker