Audio Steganography: Inaudible Audio Watermarking for Source Identification

Question

What is a relatively inaudible audio watermarking technique to overlay a broadband audio signal on top of playing music without analysing the music for low-bitrate source identification? Ideally should be resistant to multi-path, reverb and resonance.

I'm not trying to do any kind of DRM, so I don't care about resampling, time compression or frequency shifting - I just want to figure out in real-time which source is playing without adding too much audible noise to the signal.

At the moment I'm broadcasting a low bitrate ultrasonic signal on top of the music, modulated with GMSK around a carrier frequency of 19500Hz, but the high frequencies attenuate fast and the high frequencies are still audible.

I'm considering broadcasting a super wide-band PN-sequence that sounds like pink noise. I only need to sustain a transfer rate of about 3-4 bytes (24-32 bits) per second. But if I stretch out the symbol length, I'm worried about environmental noise.

I could add a delay if I need to pre-analyse the output buffer before, but I hope that I can introduce a very soft full-spectrum "hiss" or hum that can carry my source identification signal.

I don't need to track the source during periods of silence, so there will always be some music to mask my signal.

Update: I've found some prior audio steganography work, but I don't know what the trade-offs are for real-time modulation (esp. on a low-power embedded device) and if this would at all work without look-ahead on the output buffer.

Answer 1

Petrus. First of all, thanks for attaching my github library as an update into your question for other readers. I have just seen your question.

According to my own experiences, echo hiding methods are fragile towards additional echoes, so are towards reverb. A simple spread-spectrum model using a PN Sequence, which introduces some noise, embeds data into high-frequency bands, so hidden data is usually audible and not robust against low-pass filters.

Transformations such as cosine transforms and wavelet transforms are often used in audio watermarking. Applying a spread-spectrum method without a PN Sequence onto Approximation Coefficients of the third level wavelet decomposition (A3) would be very robust against many common attacks except high-pass filter, and would be almost completely imperceptible. On the other hand, hiding data onto Detail Coefficients of different levels, gives different balance between robustness & transparency. These decompositions are done very easily in Matlab if you have Wavelet Toolbox. However, I have not uploaded my related works into github yet.

We have not worked with real-time audios, but I know that all of these techniques can be applied for them as well. I am not sure whether you already solved your problem, but if you have any further questions you can point out in a comment. I am not an expert in digital signals unlike many of you on this site, so different opinions are also very welcomed.

Best Regards.

Answer 2

I'm not very familiar with the topic but I ran into this concept while working on a different problem -- its seems the use of "echo hiding" or "echo watermarking" returns a large number of results on the IEEE's site. As far as I can see, this involves adding some inaudible, unique echo signature into your signal, which can be later extracted by using the cepstrum.

One such paper, although I think this one is not a "blind" technique.