7.7 Conclusions about information hiding and watermarking

Conclusions

Different genres of music are compared (for example classical, oldies, rock, pop, hip hop, techno) with respect to their data hiding capacity and subjective sound quality when modified by each of the three encoding schemes. At peak, each one encodes over 200 seven-bit characters in a five second audio clip. Matlab becomes unstable when the algorithms are scaled up to encode more bits, although there is no indication that the algorithms would fail if given sufficient computational power. Furthermore, it is nearly impossible to distinguish between the modified and unmodified five-second sound clips, even while listening carefully through high quality headphones.

All of these algorithms can encode about the same number of bits with reasonable quality, although at least one subject complained that the PSA left a slight ringing in the marked audio file. The FMA and PSA both stand up very well to noise, while the EA does poorly. The EA quickly reaches a SNR at which the decoder can no longer tell which values were supposed to be ones and which were zeros. The FMA decoder examines such a small band of frequencies that it is much less affected by broadband Gaussian noise. Since the human ear detects phase shifts poorly we can shift the phase such that the added noise is insignificant. The FMA changes the signal power the most, while the EA changes the power the least.

In conclusion, we were successful in hiding a binary ASCII string in audio files without audible loss of quality, using three different methods: Frequency-Masking, Phase-Shifting, and Echos.