Abstract:This paper introduces a blind watermarking based on a convolutional neural network (CNN). We propose an iterative learning framework to secure robustness of watermarking. One loop of learning process consists of the following three stages: Watermark embedding, attack simulation, and weight update. We have learned a network that can detect a 1-bit message from a image sub-block. Experimental results show that this learned network is an extension of the frequency domain that is widely used in existing watermarking scheme. The proposed scheme achieved robustness against geometric and signal processing attacks with a learning time of one day.
What problem does this paper attempt to address?
The main problems that this paper attempts to solve are to improve the robustness and invisibility of digital watermarking technology, especially when facing geometric attacks and signal - processing attacks. Specifically:
1. **Robustness**: The watermark can still be correctly extracted after various attacks (such as JPEG compression, noise addition, geometric transformation, etc.).
2. **Invisibility**: The visual quality of the image after the watermark is embedded should not be significantly affected.
### Main Contributions
To solve the above problems, the author proposes a blind watermarking method based on convolutional neural network (CNN). Compared with the traditional frequency - domain - based methods (such as QDFT), this method optimizes robustness and invisibility in the following ways:
- **Iterative Learning Framework**: This framework includes three stages - watermark embedding, attack simulation, and weight update. Each stage cooperates with each other to gradually improve the model's adaptability to different attacks.
- **Adaptive Learning**: By simulating multiple attacks and adjusting the weights of the CNN, the model can adaptively respond to new attack types.
- **Surpassing Traditional Methods**: The experimental results show that the proposed CNN model is superior to the existing QDFT - based methods in terms of robustness and can effectively resist multiple attacks while ensuring invisibility.
### Mathematical Formulas
The mathematical formulas involved in the paper are as follows:
- **Gradient Descent Update Formula**:
\[
B(t + 1)=B(t)-\alpha\nabla_BL(W, B(t), m)
\]
where \(B(t)\) is the block at the \(t\)-th iteration, \(\alpha\) is the embedding rate, \(W\) is the weight of the CNN, and \(L\) is the loss function.
- **Loss Function**:
\[
L(W, B, m)=-\log p(m|W, B)+\frac{1}{2}\lambda\|B - B_0\|^2_2
\]
where \(p(m|W, B)\) is the probability that the CNN predicts the message \(m\) is embedded in the block \(B\), and \(\lambda\) is the regularization parameter.
- **Peak Signal - to - Noise Ratio (PSNR)**:
\[
\text{PSNR}(I, I') = 10\log_{10}\left(\frac{255^2\times3MN}{\|I - I'\|^2_2}\right)
\]
where \(I\) and \(I'\) are the original image and the image with the embedded watermark respectively.
- **Normalized Correlation Coefficient (NC)**:
\[
\text{NC}(w, w')=\frac{\langle w, w'\rangle}{\|w\|_2\|w'\|_2}
\]
where \(w\) and \(w'\) are the original watermark and the watermark detected after the attack respectively.
Through these improvements, this method not only improves the robustness and invisibility of the watermark but also shows its potential in practical applications.