Steganalysis of audio: attacking the Steghide

In this paper, we present a steganalytic method that can reliably detect messages hidden in WAV files using the steganographic tool Steghide. The key element of the method is mining the correlation between wavelet coefficients in a short-duration (about 20ms) in each subband. This is done by performing a four-level 1D wavelet decomposition of the audio signals, using a linear predictor for the magnitude of wavelet subband coefficients to extract significant statistics features, and employing support vector machines to detect the existence of hidden messages. Experimental results indicate that the messages embedded as small as 5% of the steganographic capacity can be reliably detected.