Abstract: Neural network steganography holds potential for copyright protection and covert communication. However, most existing methods rely on intermediate layer parameters of the model, which often leads to limited embedding capacity and insufficient robustness against noise. To address these issues, this paper proposes a steganographic scheme based on multi-level cascaded dynamic embedding. Firstly, a multi-layer weight distribution function is designed to adaptively allocate and embed secret data across multiple layers of the neural network using dynamic weights, thereby expanding the embedding capacity and enhancing robustness. Furthermore, a channel attention module is introduced into the model to counteract the performance degradation caused by secret data embedding, leveraging enhanced attention features to balance information embedding with model functionality preservation. Extensive simulation experiments demonstrate that the proposed scheme effectively addresses data hiding challenges and outperforms existing methods in terms of embedding capacity, robustness, and security.

Key words: model steganography, neural network, channel attention module, information hiding