关键词: 摘 要: 深度学习, 图像隐写, 条件对抗生成网络, 混合注意力机制, 以图藏图

Abstract: Abstract: Current deep image-in-image steganography methods face challenges in practical applications for privacy protection and covert communication due to insufficient security of stego images and distortions in recovered secret images. To address these issues, a novel image-in-image steganography framework was proposed based on conditional generative adversarial networks (cGANs) and a hybrid attention mechanism. First, a hybrid attention module was introduced into the generator network to comprehensively learn image features from both channel and spatial dimensions, enhancing the visual quality of stego images. Second, residual connec-tions were incorporated to reduce feature loss of secret images during network training, and noise-free extraction of secret images was achieved through adversarial training between the extractor and discriminator. Additionally, adversarial training between the generator and a steganalyzer was implemented to improve stego image security. Experiments conducted on public datasets (e.g., COCO) demonstrated that the proposed method outperforms existing steganography methods. Specifically, the PSNR of stego images and decrypted secret images reached 40.58 dB and 36.63 dB, representing improvements of 5 dB and 4 dB, respectively, while their SSIM values increased by 3.34% and 1.84%. In terms of security, the steganalyzer achieved an accuracy (ACC) of 0.512 (a reduction of 0.12) and a false negative rate (FNR) of 42.52% (a 27% improvement), indicating significantly enhanced security against detection.

Key words: Keywords: Deep learning, Image steganography, Conditional adversarial generation network, Hybrid attention mechanism, image into image