關於 Jsteg 的點點滴滴 (三) : T. Tan

中國模式識別國家重點實驗室(NLPR) 在 ICPR04 (23-26 Aug., Cambridge Univ., UK.) 與 ICIP04 (24-27 Oct., Singapore) 兩個研討會都有論文提出, 論文所提出的方法大致是一樣的, 不過, 實驗範圍則有所延伸。

ICPR04 : "On Estimation of Secret Message Length in JSteg-like Steganography"
ICIP04 : "Model Based Steganalysis"

第一篇論文是針對 Jsteg-like steganography 的兩種嵌入次序 (sequential or random) 做隱藏分析; 第二篇論文則是將實驗範圍延伸到 QIM (Quantization Index Modulation) 嵌入技術。

基本上, 兩篇論文在描述 Jsteg 嵌入技術的文章都是一樣的。

The frequently used steganographic method in JPEG format is the JSteg-like algorithm, which is proposed by D. Upham [I]. It works by embedding message bits as the LSBs of the quantized DCT (Discrete Cosine Transform) coefficients. The embedding mechanism skips all coefficicnts with the values of '0' or '1'. There are two embedding ways according to the selection of coefficients. One is sequential embedding; the other is random embedding whose coefficients selection is usually determined by a secret stego key shared by the communicating parties.

在這篇論文中, 還有一點我覺得很有趣！可以做實驗看看。作者引用 P. Sallee 的 IWDW03 論文 "Model based steganography" [ PDF ] 所發現的, 認為用 generalized Cauchy distribution 來描述量化後的 DCT 係數的分布會比用 generalized Laplacian/Gaussian distribution 還要來得精確。

Sallee used a specialized form of a generalized Cauchy distribution instead of the generalized Laplacian. When taking into account a more accurate estimation of the quantization effects, Sallee found this distribution appears to fit DCT coefficients better than the generalizcd Laplacian/Gaussian.