Aiming at resolving the problem of the irreversibility in some common neural networks for secret data extraction, a novel image steganography framework is proposed based on the generator of GAN (Generative Adversarial Networks) and gradient descent approximation. During data embedding, the secret data is first mapped into a stego noise vector by a specific mapping rule, and it is input into the generator of a GAN to produce a stego image. The data extraction is accomplished by iteratively updating the noise vector using the gradient descent with the generator. When the error is declined within the allowable error, the output image of the generator is approximate to the stego image, and the updated noise vector will also approach to the stego noise vector. Finally, the secret data is extracted from the updated noise vector. Experiments and analysis with WGAN-GP (Wasserstein GAN-Gradient Penalty) show that it can achieve good performance in extraction accuracy, capacity and robustness. Furthermore, the discussions also illustrate its good generalization with different GAN models and image datasets.

中文翻译：

---

基于生成对抗网络和梯度下降逼近的鲁棒无盖隐写术

针对一些常见的神经网络用于秘密数据提取的不可逆性问题，提出了一种基于GAN（Generative Adversarial Networks）生成器和梯度下降逼近的新型图像隐写框架。在数据嵌入过程中，秘密数据首先通过特定的映射规则映射到隐写噪声向量中，然后输入到 GAN 的生成器中以生成隐写图像。数据提取是通过使用生成器的梯度下降迭代更新噪声向量来完成的。当误差在允许误差范围内下降时，生成器的输出图像接近于隐秘图像，更新后的噪声向量也将接近隐秘噪声向量。最后，从更新后的噪声向量中提取秘密数据。用WGAN-GP（Wasserstein GAN-Gradient Penalty）的实验和分析表明，它可以在提取精度、容量和鲁棒性方面取得良好的表现。此外，讨论还说明了它对不同 GAN 模型和图像数据集的良好泛化。