Jeffrey Bloom

1
Understudied Constraints Imposed by Watermarking
Applications

• Jeffrey Bloom
• Dialogic Research
• (Some work done while at Thomson)

With post-presentation comments added in green.
Workshop on Multimedia, Mathematics, and Machine
Learning II Banff International Research Station
for Mathematical Innovation and Discovery Banff,
Alberta

July 5-10, 2009
2
Understudied Problems

• Watermark embedding in the compressed domain
• Watermark detection in the compressed domain

3
Compressed Domain Watermarking
Original Compressed Stream
Marked Compressed Stream
Watermark Embedding
Decode
Encode
Partial Decode
This slide discusses the evolution of
watermarking compressed streams from full
decompression to partial decompression to true
compressed domain watermarking the last of which
is the subject of the first part of this
presentation. For situations where partial
decode is too expesive.
4
Stream Replacement Watermark Embedding


Shows the concept of stream replacement
watermarking. A chunk of the stream is replaced
with different data (in brown). But where does
the brown data come from? Need to step back and
review some basic frameworks on the next few
slides.
5
Watermarking Frameworks
It is this last case (informed embedding /
informed detection) that represents the
application that motivates this work. An
informed embedder can analyze the content
off-line and generate side information for the
embedder and for the detector.
Where does the D come from?
6
2-stage Embedding
The Analyze part of the embedder can be
separated from the Embed part. The analysis can
take place in a powerful server where it can
examine the pixel data, the compressed domain
syntax elements, and the entropy encoded
bitstream.
Analysis
Entropy Decode
Full Decode
D
Analysis
E
P
E
Stream Embedding
7
Example Blu-ray
BD Player
Virtual Machine
Media Transform
Self-Protecting Digital Content
SPDC is described by CRI in their white paper
http//www.cryptography.com/resources/whitepapers/
SelfProtectingContent.pdf Use of this technology
in Blu-ray is discussed at http//www.cryptography
.com/technology/spdc/index.html among other
places.
8
SPDC

• Multiple values for each repair
• Enables forensic watermarking during repair

Damaged compressed stream is not valuable to a
pirate. Damage is done off-line, repair is done
at play time. Repair info must be carefully
protected.
9
Identifying the hard problem

• Blu-ray
• MPEG2, VC1, H.264/AVC
• H.264/AVC
• CAVLC - Context-adaptive variable-length coding
• CABAC - Context-based adaptive binary arithmetic
  coding
• VLC case has been addressed
• CABAC is hard
• Arithmetic Code
• Context Adaptive

Blu-ray supports three compression standards
VC1 is not widely used
H.264 supports two entropy encoding schemes
Both MPEG2 and CAVLC
10
Will be more widely needed
E
Stream Replacement
CABAC Encoded Bitstream
CABAC Encoded Bitstream

• This scenario is not limited to Blu-ray
• CABAC-encoded H.264 is becoming widely deployed
• High volume watermarking systems will not have
  the luxury of doing an entropy-decode/watermark/e
  ntropy-encode cycle

For adoption of watermarking in high volume
network applications, this is an important and
understudied problem
11
Mobile Video Handoff Points
What happens when I want to watch The Daily Show
on my mobile phone from here in Banff? Many
different people touch the content. This slide
shows an example of the how the video gets to my
phone.
Each participant will optimize the content for
its own network. This often involves transcoding.
Watermarks can be used to help track content.
Detectors distributed throughout multiple
networks.
12
Watermarking Challenge

• Tracking watermark embedded by content owner
• Embedding can be done in any convenient domain
• Watermark detection at various points in the
  network for tracking
• Watermark must be recoverable from any compressed
  domain without decoding
• MPEG2
• H.263
• MPEG4
• H.264/AVC

13
Watermarking Challenge

• Assume that we can do entropy decode

Transcoder
Transcoder
Encode
Decode
Encode
Decode
Encode
At any point, the stream can be decoded and the
watermark recovered from the pixels
Conceptually, we can consider a transcoder as a
decode/encode pair with pixels in the middle.

• Intermediate transcodes look like noise ? model
  as a single transcode plus noise
• Information is in there

Note the colors indicate matching encode/decode
pairs (same coding standard). Consider pixels
before the red encode and the pixels after the
red decode. The pixels after are the same as the
pixels before plus coding noise from the red
encode.
14
Watermarking Challenge

• Assume that we can do entropy decode

Encode
Decode
Encode
Decode
Encode

• Intermediate transcodes look like noise ? model
  as a single transcode plus noise
• Information is in there

15
Watermarking Challenge

• Assume that we can do entropy decode

Encode
Encode

• Intermediate transcodes look like noise ? model
  as a single transcode plus noise
• Information is in there

16
Watermarking Challenge

• Assume that we can do entropy decode

pixels
pixels noise
Transcoder
Encode
Decode
Encode

• Intermediate transcodes look like noise ? model
  as a single transcode plus noise
• Information is in there

17
Summary

• Watermark embedding in the entropy-coded
  compressed domain
• Watermark detection in the compressed domain
  after uncontrolled transcoding

For adoption of watermarking in high volume
network applications, this is an important and
understudied problem
For adoption of watermarking for tracking in
network applications, this is an important and
understudied problem