Convolutional Steganalysis Network
Definition
A deep neural network trained end-to-end to classify cover versus stego images. Architectures such as XuNet, SRNet, and Yedroudj-Net use learned high-pass filter preprocessing layers followed by convolutional feature extraction, avoiding manual feature engineering.
- Examples
- XuNet, SRNet, Yedroudj-Net
- Task
- Classify cover versus stego images
- Feature extraction
- Learned, not hand-engineered
Common questions
How does a convolutional steganalysis network differ from classical statistical steganalysis?+
Classical methods rely on hand-crafted statistical features such as pixel co-occurrence, while these networks learn discriminative features directly from labelled training data through their convolutional layers.
Why do these networks typically start with a high-pass filter preprocessing layer?+
It emphasises the faint, noise-like residuals where steganographic embedding is most detectable, suppressing the dominant image content that would otherwise overwhelm the network's learning process.
Related terms
- Adaptive Embedding
- A steganographic strategy that concentrates embedding changes in high-texture or high-noise image regions where they are perceptually and statistically harder to detect....
- Calibration
- A steganalysis technique that estimates the cover image statistics by cropping or decompressing and re-compressing the test image to produce a reference....
- Chi-Square Attack
- A specific steganalysis test for LSB substitution in images. It tests whether pairs of pixel values that are related by flipping the...
- LSB Substitution
- Least-significant-bit substitution: the lowest-order bit of each sample value (pixel colour channel or audio sample) is overwritten with one bit of the...
- Rich Model (SRM)
- A high-dimensional feature set for steganalysis constructed from joint statistics of pixel prediction residuals computed with many different filter kernels and quantisation...