Content-based audio authentication watermarking techniques extract perceptual relevant audio features, which
are robustly embedded into the audio file to protect. Manipulations of the audio file are detected on the basis of
changes between the original embedded feature information and the anew extracted features during verification.
The main challenges of content-based watermarking are on the one hand the identification of a suitable
audio feature to distinguish between content preserving and malicious manipulations. On the other hand the
development of a watermark, which is robust against content preserving modifications and able to carry the
whole authentication information. The payload requirements are significantly higher compared to transaction
watermarking or copyright protection. Finally, the watermark embedding should not influence the feature
extraction to avoid false alarms. Current systems still lack a sufficient alignment of watermarking algorithm
and feature extraction.
In previous work we developed a content-based audio authentication watermarking approach. The feature
is based on changes in DCT domain over time. A patchwork algorithm based watermark was used to embed
multiple one bit watermarks. The embedding process uses the feature domain without inflicting distortions to
the feature. The watermark payload is limited by the feature extraction, more precisely the critical bands. The
payload is inverse proportional to segment duration of the audio file segmentation. Transparency behavior was
analyzed in dependence of segment size and thus the watermark payload. At a segment duration of about 20 ms
the transparency shows an optimum (measured in units of Objective Difference Grade). Transparency and/or
robustness are fast decreased for working points beyond this area. Therefore, these working points are unsuitable
to gain further payload, needed for the embedding of the whole authentication information.
In this paper we present a hierarchical extension of the watermark method to overcome the limitations given
by the feature extraction. The approach is a recursive application of the patchwork algorithm onto its own
patches, with a modified patch selection to ensure a better signal to noise ratio for the watermark embedding.
The robustness evaluation was done by compression (mp3, ogg, aac), normalization, and several attacks of the
stirmark benchmark for audio suite. Compared on the base of same payload and transparency the hierarchical
approach shows improved robustness.
|