Abstract: Researching on the auto-correlation in steganographic software tools Hide4PGP，a group of new feature is proposed which is implemented by using support vector machine to classify with gradient energy variation and linear prediction statistics variation features after wavelet decomposition.The variations root in audio carriers with and without messages，and carriers already contain messages and re-embedded by Hide4PGP.This scheme can detect the existence of stego-audio embedded by Hide4PGP from clear-audio effectively.On the premise of high detection rate，an improvement of features by H.Farid has been made with a lower feature dimension and shorter classified time.

Key words: audio steganalysis, re-embedding, wavelet decomposition, linear prediction coding

摘要： 针对隐写软件Hide4PGP隐写自相关性进行研究，以隐写一次与未隐写音频和隐写一次与隐写两次音频梯度能量的变化量，和将上述音频样本三级小波分解后，各级小波域细节分量线性预测误差的统计量（均值、方差、偏度和峰度）的变化量为特征向量，用支持向量机作分类器进行音频隐写分析。实验表明，这组特征可以有效地检测出经隐写软件Hide4PGP隐写的音频文件。在保持有效的检测率前提下，降低了特征维数，缩短了SVM分类时间，是对H.Farid所提特征的改进。

关键词: 音频隐写分析, 二次嵌入, 小波分解, 线性预测