摘要
自适应图像隐写算法是一种以图像为载体,通过手工设计嵌入失真代价,指导隐写码在图像载体中嵌入秘密消息的信息隐藏算法.长期以来,这类算法将秘密消息尽可能隐藏在图像纹理更深更复杂的位置以对抗基于富特征的隐写分析检测.然而,伴随着深度学习在隐写分析领域的快速发展,人工设计的自适应算法受到严重挑战.此外,基于加性失真的隐写编码在嵌入消息时,复杂纹理向边界聚集所产生的统计异常问题也亟待解决.因此,本文总结了各类人工失真代价的优势和不足,归纳出当前自适应算法在空域的设计范式,并结合UNIWARD在各嵌入域的转换规则,提出基于嵌入失真代价ρ的通用域隐写转换公式.然后,从隐写嵌入失真代价与图像纹理稀疏关系的角度出发,以Canny算子划分纹理、Gauss模糊缩放轮廓、AutoML搜索阈值的方式,提出了一种通用域隐写算法Canny Gauss.实验结果表明,本文所提通用域隐写转换公式能够有效应用于现有主流算法.同时,在UNIWARD所有可行嵌入域中,本文所提算法表达出更高嵌入失真代价稳定性和隐写隐蔽性,在第三方权重加持下的深度隐写分析表现与UNIWARD相比至少提升2.6%、最高提升14.6%.这为自适应隐写算法的通用域设计,以及抵抗基于纹理特征的深度隐写分析检测提供了新思路.
Adaptive image steganography algorithms have emerged as a means of concealing secret messages within image carriers,employing a manual design of distortion costs to guide the process of message embedding.The primary objective of these algorithms has been to hide secret information in regions of the image that possess intricate and complex textures,thereby thwarting feature-based steganalysis detection methods.However,the rapid advancements in deep learning within the field of steganalysis have posed significant challenges to the efficacy of manually designed adaptive algorithms.Furthermore,there is a pressing need to address the statistical anomalies that arise from the aggregation of complex textures towards the boundaries when employing additive distortion-based steganographic encoding techniques.To tackle these challenges,this paper provides a general summary of the strengths and limitations associated with various handcraft distortion cost design.It also presents an paradigm of the existing design paradigms for adaptive algorithms in the spatial domain,considering the transformation rules of UNIWARD across different embedding domains.In order to improve upon the existing techniques,the paper proposes a universal domain steganographic transformation formula based on the embedding distortion cost p.This formula provides a flexible framework that can be applied to a wide range of mainstream algorithms,enhancing their performance and adaptability.Moreover,this paper introduces a groundbreaking universal domain steganographic algorithm known as Canny Gauss,which capitalizes on multiple techniques to achieve highly effective message embedding.Firstly,the algorithm employs the Canny operator to perform texture segmentation,enabling the identifi-cation and selection of regions within the image that possess rich texture information suitable for embedding secret messages.By leveraging this approach,Canny Gauss ensures that the embedded messages are strategically placed within areas that can effectively camouflage the hidden information.In addition,the algorithm utilizes Gaussian blur to scale the contours of the image.This step is crucial in guaranteeing a seamless integration of the embedded messages with the surrounding textures,making them inconspicuous to visual inspection and steganalysis techniques.To further optimize the performance of the algorithm,an AutoML framework is employed to automatically search for suitable threshold values.This technique enhances the overall robustness and effectiveness of the steganographic process by dynamically adjusting the thresholds based on the characteristics of the input image.By adapting the thresholds to each specific image,Canny Gauss maximizes the concealment of secret messages while minimizing any adverse effects on image quality or detectability.Experimental results demonstrate the efficacy of the proposed universal domain steganographic transformation formula when applied to existing algorithms.In comparison to UNIWARD,the algorithm presented in this paper exhibits improved stability in embedding distortion costs and enhanced steganographic security.Moreover,when coupled with third-party weights,the algorithm showcases notable improvements in deep steganalysis performance,with a minimum enhancement of 2.6%and a maximum enhancement of 14.6%compared to UNIWARD.This paper not only provides valuable insights for the design of adaptive steganography algorithms in universal domains but also offers a new strategie to counter deep steganalysis detection techniques that rely on texture features.
作者
李季瑀
付章杰
王帆
LI Ji-Yu;FU Zhang-Jie;WANG Fan(Engineering Research Center of Digital Forensics,Ministry of Education,Nanjing University of Information Science and Technology,Nanjing 210044;State Key Laboratory of Integrated Services Networks,Xidian University,Xi'an 710126)
出处
《计算机学报》
EI
CAS
CSCD
北大核心
2024年第1期213-230,共18页
Chinese Journal of Computers
基金
国家重点研发计划(2021YFB2700900)
国家自然科学基金(U22B2062,62172232)
江苏省杰出青年基金(BK20200039)
江苏省大气环境与装备技术协同创新中心基金资助。
