This work inspired by a recent work of Mo Chen et al
In fact, the algorithm proposes a simple CNN architecture that was improved using of catalyst kernels as initialization to neurons and weight propagation via transfer learning.
The tables below show the results according to the detection error of each algorithm (WOW, HUGO, S-UNIWARD) with the payloads (1.0, 0.7 , 0.5 , 0.3) bpp.
- Experiment results of
S-UNIWARD:
| payload (bpp) | 1.0 | 0.7 | 0.5 | 0.3 |
|---|---|---|---|---|
| Pe | 0.04 | 0.07 | 0.12 | 0.27 |
- Experiment results of
WOW:
| payload (bpp) | 1.0 | 0.7 | 0.5 | 0.3 |
|---|---|---|---|---|
| Pe | 0.04 | 0.08 | 0.17 | 0.33 |
- Experiment results of
HUGO:
| payload (bpp) | 1.0 | 0.7 | 0.5 | 0.3 |
|---|---|---|---|---|
| Pe | 0.04 | 0.09 | 0.16 | 0.31 |
The paper of the article will be available soon.
- Tebsorflow v > 1.0
- Keras v > 2.0
- Sickit-Learn
- OpenCV 3
Please cite our article if this code helps you in some way
El Beji R., Saidi M., Hermassi H., Rhouma R. (2018) An Improved CNN Steganalysis Architecture Based on “Catalyst Kernels” and Transfer Learning. In: Bach Tobji M., Jallouli R., Koubaa Y., Nijholt A. (eds) Digital Economy. Emerging Technologies and Business Innovation. ICDEc 2018. Lecture Notes in Business Information Processing, vol 325. Springer, Cham
MIT