Audio and Video Watermarking

1
Audio and Video Watermarking
Mr. Pirate
Joseph Huang Weechoon Teo
2
What is watermarking?
Permanent proof of originality for paper
media. Permanent proof of ownership for digital
media. Watermarking preserves intellectual
property unlike encryption. Watermarking is
statistically and physically invisible
(PRN). Watermarking can be detected even after
distortions. Watermarking is done in the
frequency, temporal, and/or spatial domains.
3
Audio Watermarking

• Robustness
• Watermark has to be robust to signal
  manipulation.
• Impossible to remove without significant
  alteration of the signal.
• Statistically undetectable by others to prevent
  the efforts of unauthorized removal.
• Can be fulfilled if the potential number of
  keys that produce distinct watermarks is large.
• Detection scheme should be as statistically
  reliable as possible.
• False rejection or acceptance of watermark
  should be minimal.

4
Audio Watermarking A Temporal Method, p.1

• Does not require original signal for the
  detection of watermark.
• Requires only a seed or key.
• Watermark is embedded into the audio signal by
  changing the least significant bits of the 16-bit
  or 8-bit audio samples.
• Results only in slight amplitude modification
  in the time domain.

5
Audio Watermarking A Temporal Method, p.2

• Watermarked signal is formed by the following
  equation

y(i) is the watermarked audio signal. x(i) is the
original audio signal. w(i) is form from a random
number generator. f(x(i), w(i)) is a function
that accounts for the basic audio masking
properties.

• S is defined as follows

6
Audio Watermarking A Temporal Method, p.3

• The watermark detection value, r, is calculated
  by the equation below

• Theoretically r 0, 1, but due to estimation
  of x(i) by y(i), r 0-e, 1e.
• A detection threshold of 0.5 can be used to
  decide on the existence of audio watermark.
• Figure on right shows the pdf for the value of
  r in a non-watermarked and watermarked signal.
  Both distributions have been calculated using
  1000 different watermarks with SNR 26.

7
Audio Watermarking Results, p.1

• Detection values in a watermarked signal using
  various seed (key is 444).
• Only the correct key yields a value of r higher
  than threshold.

• No significant shift in PDF after resampling
  from 44.1KHz to 11.025KHz and back
• 100 success in watermark detection after
  resampling

• Requantization from 16-bit to 8-bit and back
  results in increase of deviation of PDF.
• Still achieve 99.8 accuracy in watermark
  detection.

8
Audio Watermarking Results, p.2

• Filtered by a moving average filter of length
  20 which introduces a noticeable audio distortion
• Shift in mean and variance but still results in
  100 detection.

• 44.1 kHz signal Low-pass filtered by a 25th
  order Hamming LPF with cut-off at 22.05KHz.
• Shift in mean and variance but still results in
  100 detection.

• No significant shift in PDF after MPEG3 Layer
  III 80kbs lossy compression.
• Based on 0.5 threshold, still achieve 100
  watermark detection.

9
Video Watermarking

• Issues on identical watermarks for each frame
• Problems in maintaining statistical
  invisibility.
• Issues on independent watermarks for each frame
• Problems in easy removal of watermarks.
• Robustness
• Must survive frame averaging, frame dropping,
  frame swapping, cropping, temporal rescaling.
• Must be able to discern imposter watermarks
  (deadlock). Problems in use of the original video
  sequence. Problems when no video sequence is
  needed.

10
Video Watermarking Deadlocking

• Detection and Generation of Pseudorandom Sequence
• Original sequence is present for comparisons,
  but what about imposters?
• Possible solution Public/Private Key
  Pseudorandom Generator
• Embedded Watermark for added authorization

PRN
Supplied by author
11
Video Watermarking A Method, p.1
Temporal Wavelet Transform yields 1) Low-pass
frames (Static, non-moving component) 2)
High-pass frames (Dynamic, moving
component) Frequency and Spatial Masking are
tuned to human visual perception.
Spatial Masking
Temporal WT
Extract Blocks
Frequency Masking
DCT
X
IDCT
X
DCT
Video Frames
Wavelet Frames
Author signature

Watermark block
12
Video Watermarking A Method, p.2

• Detection of Watermark
• With knowledge of location in video sequence
• X input, R received coeffs, F original
  coeffs, N noise, W watermark
• H0 Xk Rk - Fk Nk (No watermark)
• H1 Xk Rk - Fk Wk Nk (Watermark)
• Without knowledge of location in video sequence
  (just one frame present)
• Only look at the low-pass frames (static,
  non-moving component)
• Decision thresholds are determined by a scalar
  similarity

Typical results
13
References

• P. Bassia and I. Pitas, Robust Audio
  Watermarking in the Time Domain. Dept. of
  Informatics, University of Thessaloniki.
• Jian Zhao, Look, Its Not There. BYTE Magazine
  - January 1997.
• M. Swanson, B. Zhu, and A. Twefik,
  Multiresolution Scene-Based Video Watermarking
  Using Perceptual Models. IEEE Journal on
  Selected Areas in Communications, IEEE 1998.

14
Answer to Questions, p.1
How is the key embedded into the watermarked
signal, y(i)?
The figure on the right shows how y(i) is
generated.

• The key is a unique code for an authors
  identification. This unique code is used to
  generate a maximum length Pseudo-random Noise
  sequence. This PN sequence is then used to
  generate the watermark signal w(i) as show in the
  diagram above. Thus the key is really utilized by
  the function w(i).
• A masking threshold for the audio signal can be
  generated using MPEG Audio Psychoacoustic Model
  1. The PN sequence generated by the key is then
  filtered with the masking filter M(w) to ensure
  that the spectrum of the watermark is below the
  masking threshold. This ensures that the
  watermark is inaudible after embedding into the
  signal.

15
Answer to Questions, p.2

• What does statistically undetectable mean?
• How do we make a watermark statistically
  undetectable?
• By statistically undetectable, we mean that a
  pirate is unable to detect the watermark simply
  by generating the whole set of all possible
  watermarks. In other word the possibility of a
  pirate correctly guessing the right key is close
  to zero. This is to ensure that a pirate is
  unable to remove or claim ownership for the
  watermark in the audio signal.
• The condition for statistically undetectable
  is simply fulfilled by having a huge set of keys
  that will generate distinct watermarks. This will
  result in statistical safety for the watermarked
  audio signal.