Blind detection on echo hiding based on cepstrums

An effective blind detection for four kinds of echo hiding based on three cepstrums is proposed. Due to statistic characteristic changing with echo in stego speech, real cepstrum, complex cepstrum and power cepstrum are analyzed. Firstly we calculate high order statistics of each cepstrum, including coefficients of skewness and kurtosis in combination. Then we train the combinatorial skewness and kurtosis in each cepstrum of orignal speech and echoed speech with SVM. By obtaining combinatorial skewness-kurtosis thresholds, we can used them for blind steganalysis of echo hiding. Experimental results show that our proposed blind detection with three cepstrums performs well in testing Ker.1 and Ker.3 echo hiding, and power cepstrum achieves high accuracy in four kinds of echo hiding.