[LAU] Looking for Audio Watermarking Advice & Tools

[Fons Adriaensen]

On Sun, Oct 23, 2011 at 10:09:44PM -0400, Jeremy Salwen wrote:

> Hmm... it seems to me that if the watermark does not cause audible
> distortion, then shouldn't there be a simple algorithm to remove the
> watermark: namely, to watermark the same file again with different
> information?

Not always. For example, a spread-spectrum signal can co-exist
with others if it uses a different spreading code. And even if
the algorithm is know, the actual code used can be kept secret.

> I suppose this requires you to know the algorithm used to watermark the file
> in the first place.  All the algorithms discussed here:
> http://www.ece.uvic.ca/~aupward/w/watermarking.htm seem like they would be
> pretty trivially removed by the applying them again with different watermark
> data.

See above. And there are *much* more sophisticated ways than those
discussed in that page.

There are three steps involved in adding a watermark to an audio
file, and they are similar to those found in any digital telecom
system:

1. Encryption,
2. Coding - inserting redundancy to allow error recovery,
3. Modulation - representing the digital data in audio form,

and of course the inverse sequence to recover a watermark:

3. Demodulation - extracting features from the audio signal
   to produce 'soft bits' (bits which are not hard 0 or 1
   but are represented by a probability function),
2. Decoding - combining soft bits to produce 'hard' ones,
1. Decryption.

As long as these steps are separate, having the source code
reveals all about 2 and 3, and that is usually enough to
allow removing a watermark without introducing significant
damage to the audio. 

In digital telecoms all 'advanced' systems combine 2 and 3
to some extent, mainly to improve both bandwidth and power 
efficiency. There's been a lot of research into such this,
and the times when e.g. 16-QPSK + Viterbi + Reed-Solomon
was 'state-of-the-art' are well past.

What is required for a robust watermarking scheme is to 
mix in (1) as well, and one way is using cryptographically
strong spreading codes.  
 
Ciao,

-- 
FA