A Robust Image Steganography Framework Against Unknown Attacks Using Adversarial Attack Generation and Dynamic Mixture of Experts

Saideh Naderi¹ Reza Gholamrezaei²

1) Ph.D. Student, Department of Computer Engineering, Kerman Branch, Islamic Azad University, Kerman, Iran Email:
2) Department of Computer Engineering, Kerman Branch, Islamic Azad University, Kerman, Iran Email:

IOI Code : AAHD-ABHBHA
Publication : National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
    Deep learning-based image steganography has emerged as a prominent approach in information security due to its high embedding capacity and the satisfactory visual quality of cover images. However, most existing methods exhibit limited robustness against unknown attacks and unforeseen distortions, as they are typically trained on a restricted set of predefined attack scenarios. In this paper, a novel framework for robust image steganography against unknown attacks is proposed by integrating an adversarial attack generator with a Dynamic Mixture-of-Experts (DMoE) architecture. First, an adversarial attack generator is designed to produce a diverse range of distortions and simulate realistic communication channel conditions, thereby enhancing the model’s generalization capability. Second, instead of relying on a single decoder, multiple specialized decoders are developed to handle different categories of attacks. An intelligent routing mechanism dynamically selects the most appropriate expert, or an optimal combination of experts, according to the characteristics of the received image. This architecture enables adaptive and robust decoding under varying and previously unseen attack conditions. The performance of the proposed framework will be evaluated using message recovery rate, bit error rate (BER), visual quality metrics, and structural similarity measures. It is expected that the proposed method will significantly improve robustness and information recovery performance against unknown attacks while preserving high visual quality, outperforming conventional image steganography approaches.

Keywords : image steganography deep learning unknown attacks adversarial attack generator dynamic mixture of experts information security