目的 为了保证载密图像的抗统计分析能力同时避免对特定载体模型的过度优化，提出一种以最小化失真为目标的隐写算法。方法 算法以各方向元素基团为基本单元定义失真函数，以Fisher准则函数的极大值为标准对失真参数进行优化，将失真函数与统计特征相关联。在秘密信息的嵌入过程中，首先依据邻域系统将图像载体分为若干元素阵列，令不同的阵列对应不同的特征子集，再利用Gibbs抽样和STC(Syndrome-trellis code)编码实现对这些有所差异的特征子集的集成，从而在最小化失真的同时保持载体的统计特征。结果 在3组不同维数的检测算法下比较该算法与同类算法的分类误差。结果表明，该算法能更好地保持统计模型，嵌入率为0.5 bit/pixle时相应特征集的检测误差仍高于0.4，面临高维检测时算法同样具有较高的安全性。结论 该算法借助最小失真思想实现了隐写前后统计特征的保持，且有效避免了在不完整模型上的过度优化，拥有比同类算法更好的适应性和安全性。

Objective The development of steganalysis technology, especially universal blind steganalysis based on statistical characteristics, has increased the demand for steganographic security against statistical analysis. To ensure stego-image security against statistical analysis while avoiding overtraining to an incomplete cover model, this study presents a steganographic Method that minimizes embedding distortion.Method A novel distortion function reflecting higher order statistics is first defined based on element cliques. According to the Results of theoretical derivation and experiments, the maximal value of the Fisher criterion function is used as the optimization criterion for the parameters in the distortion function, such that the distortion function can be related to statistical detectability. Finally, when a secret message is embedded, multiple different feature subsets are integrated through Gibbs sampling and syndrome-trellis coding.Statistical characteristics are preserved while distortion is minimizes.Result Experiments are proposed to compare the classification errors of the new Method with those of three similar steganalysis Methods with different dimensions. Results show that the new Method can better preserve image model and maintains high security even when detected using a high-dimensional steganalysis Method. The classification error using corresponding feature set is higher than 0.4 while the embedding rate is 0.5 bit/pixel.Conclusion The new steganographic Method successfully preserves statistical characteristics while minimizing distortion function. Moreover, the proposed steganographic Method effectively avoids overtraining to an incomplete model and has better adaptability and security than similar Methods.