R/D optimal data hiding

State of the art audio coders exploit the redundancy in audio signals by shaping their quantization noise below the signal's masking curve, which is a signal-dependent threshold of audibility. This framework can be extended to the context of data hiding, where the data play the role of noise. To minimize audio distortion the data power should be closely adapted to the time-varying masking curve; each power switch, however, reduces the net throughput via its associated side information. This tradeoo can be cast in a rate/distortion framework: the optimal sequence of power levels and the optimal sequence of power switchpoints is found by minimizing a Lagrange cost functional relating perceptual audio distortion to throughput, and is implemented as a linear-time trellis search. For 16-bit, 44.1 KHz PCM stereo signals, a net throughput of the order of 30 kbits/sec can usually be achieved at no perceptual cost in an algorithmically eecient way.