Abstract: Addressing the challenge of balancing imperceptibility and robustness in information hiding, a robust image steganography model guided by an attention mechanism and a modified dense convolutional network is proposed based on a generative adversarial network (GAN). Firstly, within the encoder, feature extraction is performed using the modified dense convolutional network. Shallow and deep features of the image are fused with the secret information through dense connections, thereby enhancing the model's robustness. Secondly, to minimize distortion caused by information hiding, a squeeze-and-excitation (SE) channel attention mechanism is introduced to automatically select texture-rich regions for embedding. Finally, the quality of the images generated by the encoder is enhanced through adversarial training between the encoder and the decoder. Experimental results demonstrate that, compared to other methods, the proposed model achieves a peak signal-to-noise ratio (PSNR) of 51.7955 dB, a structural similarity index (SSIM) of 0.9990, and a bit error rate (BER) of 0.0001 on the COCO2017 dataset. These results are observed to outperform those of other methods, indicating that the proposed model effectively balances imperceptibility and robustness during the information hiding process.