Adversarially Robust Mask Generator: A Secure Encoder Network for Deep Learning-Based Steganography

We propose the Adversarially Robust Mask Generator (ARMG), a novel encoder network for deep learning-based steganography that simultaneously achieves high embedding fidelity and certifiable security against steganalytic attacks. Traditional steganographic methods often suffer from detectable artifacts or vulnerability to adversarial perturbations, hence limiting their practical deployment. The ARMG addresses these challenges by integrating a U-Net-style mask generator with adversarial training, gradient masking, and Lipschitz-bound certification into a unified framework. The mask generator produces pixel-wise perturbations constrained to preserve visual quality while embedding secret data, whereas a Vision Transformer-based discriminator adversarially trains the system to evade detection. Moreover, the inclusion of a certifiable robustness module ensures stability against input perturbations, providing formal security guarantees absent in prior GAN-based approaches. The proposed method employs residual dense blocks with channel attention for high-capacity embedding and introduces non-differentiable quantization to obfuscate gradients during white-box attacks. Experimental validation demonstrates that ARMG outperforms existing methods in both undetectability and robustness, achieving state-of-the-art performance across multiple steganalytic benchmarks. This work bridges the gap between adversarial robustness and steganographic security, offering a principled solution for real-world applications where both data hiding and resistance to analysis are critical.