网络与信息安全学报2025，Vol.11Issue(5)：126-136,11.DOI:10.11959/j.issn.2096-109x.2025053
ISEM:基于生成多对抗网络的图像隐写增强模型
ISEM:image steganography enhancement model based on generative multi-adversarial network
摘要
Abstract
Image steganography techniques were typically categorized into symmetric and asymmetric types based on whether the embedding costs of+1 and-1 within the same embedding unit were equal.The interrelationship be-tween adjacent pixels was not fully utilized by symmetric steganography techniques,which limited the enhance-ment of steganographic security.Although asymmetric steganography techniques were able to significantly im-prove security,existing methods were often faced with issues such as unstable training processes,reliance on manual experience,or dependence on preset parameters when generating asymmetric embedding costs,making it difficult to ensure steganographic security.To address these issues,an image steganography enhancement model(ISEM)based on a generative multi-adversarial network was proposed.A relatively lightweight dual-branch struc-ture was adopted by the generator network,while the discriminator network was composed of three deep learning-based steganalysis networks.The training of ISEM was divided into two stages.In the first stage,the mapping rela-tionship between the carrier image and existing embedding costs was learned.In the second stage,the initially learned embedding costs were adjusted through adversarial training to generate high-security asymmetric embed-ding costs.Additionally,a geometric mean loss was introduced in the generator,and a mean absolute deviation loss was incorporated in the discriminator to further optimize the training effect.Experimental results show that the se-curity of both symmetric and asymmetric steganography methods are effectively enhanced by ISEM.Compared to existing asymmetric steganography methods,the highest average security is achieved by the proposed approach.关键词
图像隐写/生成对抗网络/几何均值损失/均值绝对偏差损失Key words
image steganography/generative adversarial network/geometric mean loss/mean absolute deviation loss分类
信息技术与安全科学引用本文复制引用
周赛星,叶苗欣,骆伟祺..ISEM:基于生成多对抗网络的图像隐写增强模型[J].网络与信息安全学报,2025,11(5):126-136,11.基金项目
国家自然科学基金(62472458) （62472458）
广东省自然科学基金(2025A1515012823) The National Natural Science Foundation of China(62472458),The Natural Science Foundation of Guan-dong Province(2025A1515012823) （2025A1515012823）
