Abstract: An audio steganalysis method for echo hiding based on statistical features of power cepstrums was proposed. In this scheme, audio signals were first divided into little segments and a hanning window was applied to each segment, then average power cepstrums of the windowed audio segments were calculated, and based on the feature that the average power cepstrums of echo hiding signals can generate peaks at echo delays, variances of differences and flatness of the average power cepstrums were calculated as statistical features, and, finally, support vector machine (SVM) was implemented as a classifier. This steganalysis method can not only detect the basic single echo kernel, but also be applied to the improved echo kernels. Experimental results show that the performance of the proposed method is satisfactory, achieving high classification accuracy even with low attenuation coefficients and regardless of the length of the embedded segments of audio signals.

Key words: echo hiding, audio steganalysis, echo kernel, power cepstrum, support vector machine