Using High-Dimensional Image Models to Perform Highly Undetectable Steganography

Steganalysis of Content-Adaptive Steganography in Spatial Domain

“…However, unlike G-SVM, the ensemble classifier is capable of handling the high-dimensional HOLMES features which resulted in a consistently lower detection error P E than the error for the CDF trained with a G-SVM. HUGO is confirmed to be more secure than non-adaptive ±1 embedding but the difference is less pronounced than what was reported in [14].…”

“…The first set (MARKOV) is a direct equivalent of the second-order SPAM [13] with two differences -the first-order differences were replaced with second-order differences and the transitional probability matrix with the joint matrix (co-occurrence). It is rather interesting that by changing a single line of code SPAM turns into a significantly more powerful feature set -P E has dropped from 42% [14] to 28.6%. 2 The second row of the table informs us that the detection is even better with the MINMAX residual, while the its quantized version shaves another 1% from P E .…”

“…A recent example of this type of steganography is HUGO [14]. Although this algorithm was designed for images in raster formats, the ideas can be applied to other domains and other media types.…”

“…What distinguishes HUGO from other algorithms is that it approximately preserves a very high-dimensional feature vector and thus takes into consideration a large number of complex dependencies among neighboring pixels. With the help of advanced syndrome-coding techniques, HUGO embedding was reported undetectable using state-of-the-art steganalyzers even at rather large payloads [14].…”

“…The HUGO algorithm is described in [14] and a brief description also appears in [1] in this volume. In the next section, we introduce HOLMES -a strategy for constructing a large number of diverse features capable of detecting embedding changes in more complex parts of images.…”