1- Yedroudj-Net is a convolutional neural network implemented using the Caffe toolbox and running on a GPU card. It is composed of 7 blocks, a pre-processing block, five convolutional blocks, and a fully connected block made of three fully connected layers, followed by a softmax (see Figure).
2- Yedroudj-Net network is designed for steganalysis purpose in a 2 classes scenario (Cover or Stego). The prelimilarly results show better steganalysis performances compared to the state-of-the-art.
See the following paper for more background:
-
Paper: Yedrouj-Net: An efficient CNN for spatial steganalysis
-
Authors: Mehdi Yedroudj, Frédéric Comby, and Marc Chaumont
Copyright licence:
This work is provided by the copyright holder under the terms of Licence Creative Commons Attribution - No Commercial Use - Share under the same Conditions 4.0 International.
Graph
Overall architecture: Input image Size: 1 x 256 x 256
Layer 0: Convolution with 30 filters, size 5×5, stride 1, padding 2 Size: 30 x 256 x 256
Layer 1: Convolution with 30 filters, size 5×5, stride 1, padding 2 Size: 30 x 256 x 256 30 depth because 1 set denotes 1 filter and there are 30 filters
Layer 2: Convolution with 30 filters, size 5×5, stride 1, padding 2 Size: 30 x 256 x 256
Layer 3: Average-Pooling with 5×5 filter, stride 2 Size: 30 x 128 x 128
Layer 4: Convolution with 64 filters, size 3×3, stride 1, padding 1 Size: 32 x 128 x 128
Layer 5: Average-Pooling with 5×5 filter, stride 2 Size: 64 x 64 x 64
Layer 6: Convolution with 128 filters, size 3×3, stride 1, padding 1 Size: 64 x 64 x 64
Layer 7: Average-Pooling with 5×5 filter, stride 2 Size: 128 x 32 x 32
Layer 8: Convolution with 256 filters, size 3×3, stride 1, padding 1 Size: 128 x 32 x 32
Layer 9: Global-Average-Pooling with 32×32 filter, stride 1 Size: 128 x 1 x 1 Activation function and Batch_Normalization are used through all blocks
HOW TO CITE this work:
[Mehdi Yedroudj, Frédéric Comby, and Marc Chaumont, " Yedrouj-Net: An efficient CNN for spatial steganalysis ", IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'2018, 15–20 April 2018, Calgary, Alberta, Canada, 5 pages]
The SRM part is available on http://dde.binghamton.edu/download/.
The trained model was obtained while training the Network against WOW steganography algorithm, using a payload of 0.4 bpp.