AFRL Presentation

1
Lossless Watermarking of Entropy Coded Sources
Electrical and Computer Engineering
Department Villanova University
2
Outline

• Goals and motivation for work
• Brief overview of relevant literature
• Theory
• Codespace
• Codeword-pairing
• Binary Tree Structure
• Practiceapplying theory to JPEG
• Introduction to JPEG
• Redundancy in JPEG images
• Actual encoding and decoding
• Results
• Conclusions and Future Related Work

3
Goals

• Develop watermarking scheme for metadata
  embedding that is
• Applied directly in compressed domain
• Losslessly reversible
• File-size preserving
• Format compliant

4
Motivations

• Watermarking within the compressed domain allows
  for fast, real time applications.
• Watermarking within the VLC portion of compressed
  data will not change the existing formatallowing
  for standard software to read data while
  watermarked.
• Lossless recovery of original data allows for
  many metadata applications.

5
Previous Work

• Some algorithms embed in DCT coefficients.
• Algorithms such as JSTEG, OUTGUESS, and F5.
• Embedding in DCT coefficients requires at least
  partial decompression of data.

6
Previous Work

• Few algorithms work directly with VLCs in the
  compressed domain.
• Langelaar et al. proposed a form of LSB
  watermarking.
• Chun-Shien et al. proposed modulating DCT
  coefficient level values.
• However, both of these methods are lossy.

7
Algorithm
Watermark
Offline data analysis
VLC analysis
Compressed format data
Watermarked data

• Offline analysis may only need to occur once,
  even for different data.
• Not necessary to have full video, only VLC table
  is required for analysis.

8
Theory Outline

• Concept of codespace in relation to entropy codes
• Improved capacity through codeword-pairing
• Efficiency of binary code trees

9
Variable Length Encoding

• Variable length encoding attempts to minimize the
  average codeword length by assigning shorter
  codewords to symbols that appear more often
  within a given data stream.
• This requires a priori knowledge of the data (or
  reasonable expected distribution).
• Huffman coding is the most common method of
  variable length encoding and promises the closest
  results to entropy coding.

10
RVLCs

• VLCs are instantaneously decodable from left to
  right.
• One code can not be the prefix of another.
• i.e. 10 and 101 can not be in the same VLC table.
• Reversible variable length codes (RVLCs) are
  two-way decodable.
• Most basic RVLC codes are symmetric.
• 00, 010, 101, 0110, etc.

11
RVLCs for English Alphabet
12
Codeword Pairs

• Examine a fixed length code example
• Code consists of 10 and 11
• Codespace for this code takes up 50 of total
  possible codespace when considering data only one
  codeword long.
• Consider codeword pairs (i.e. 1010, 1011,
  1110, 1111) now only 25 of possible
  codespace is being employed.

13
Concept
codeword-pairing
VLCs
Vi
Vj
Vij
codeword-pair
RVLCs Codeword-pairs

• Codeword-pairing creates additional watermark
  capacity.

14
Detecting the Watermark
watermark
Vij
Vij
watermark
Collision Vij Vxy
Good Table

• A collision occurs when a watermarked VLC
  violates the prefix condition or causes another
  VLC to violate the prefix condition.

15
Reversing the Watermark
Initial Table
Final Table
Collision

• Multiple watermarked codeword-pairs may overlap,
  making it impossible to identify the original
  codeword-pair.
• The final table eliminates all such ambiguities.

16
Overview of Offline Analysis
Create exhaustive Pairing of codewords
Input VLC table
Locate redundant bits
Watermark bits must be unambiguous

• Limit number of bits before error must be
  discovered.
• Create codeword pairs.
• Locate redundant bits.
• Specify which redundant bits can unambiguously be
  watermared.

17
Estimated Capacity

• Estimated capacity is based on the RVLC table
  actual capacity will vary based on the compressed
  data.
• Estimated capacity is calculated by summing over
  all codeword-pairs, the product of each
  codeword-pairs probability of occurrence and
  divide by its length. The result is a percent.

18
Results

• The algorithm was applied to the encoding of the
  English alphabet using an asymmetric RVLC 3.

19
Results

• Algorithm fulfills the following criteria
• Watermark within the compressed domain.
• Does not change format of data.
• Losslessly removable.
• Algorithm has great potential for embedding
  metadata in a wide range of applications.

20
Binary Tree Structure

• Previous work
• Used computationally expensive, exhaustive
  searches to determine watermark bit locations.
• Resulted in lookup tables.
• Binary Tree Structure
• Exponentially decreases complexity for
  determining watermark bit locations.
• Result is binary tree that can be used for both
  watermarking and decoding.

21
Example Binary Tree
0
1
Leaf node
Branch node
Available node
00
010
011
100
101
110
111
0110
22
Codeword-pair Binary Tree
Leaf node
Branch node
0
1
0000
00010
01000
011000
000110
010010
0100110
0110010
01100110
23
Failed Watermark Attempt
Leaf node
Branch node
Collision node
0
1
0001
0000
00010
01000
011000
000110
010010
0100110
0110010
01100110
24
Successful Watermark Attempt
Leaf node
Branch node
Watermark node
0
1
0010
0000
00010
01000
011000
000110
010010
0100110
0110010
01100110
25
Application JPEG

• Knowledge of watermarking and binary tree
  structure applied to watermarking of JPEG images.
• Goals of JPEG watermarking
• Designed for metadata applications
• Algorithm should still be applied in compressed
  domain, file-size preserving, and lossless

26
JPEG Compression
Each block is forward DCT transformed
Quantized coefficients are zigzag scanned into
one-dimensional array
Raw image
Add headers and markers JPEG file
Group into 8x8 pixel blocks
Each block is quantized
Entropy encoded

• Quantization table is 8x8 matrix
• Increasing values in quantization table decreases
  file size, but deteriorates image quality?allows
  for various compression rates
• Entropy encoding can be Huffman or arithmetic

27
Redundancy Custom Vs. Standard
?Example AC VLC table has 162 codewords. ?Actual
images typically require less than half.
Custom
Standard

• JPEG Standard allows for custom AC VLC tables to
  be created for each image
• JPEG Standard also has an example table that
  includes every possible run/size combination

28
Redundancy

• Many images use the example AC VLC table provided
  in the standard regardless of content in the
  image.
• Popular software tools such as MATLAB and
  Microsoft Paint use the example table when
  creating JPEG images.
• Since AC VLC table is not optimized for specific
  image, the entropy coding is not perfectmeaning
  that there is some inherent redundancy.

29
Example Binary Tree
0
1
Leaf node
Branch node
Available node
00
010
011
100
101
110
111
0110
30
Watermarking
Binary Tree Find Watermark bit locations in
VLCs
JPEG image
Watermarked JPEG Image
Check capacity of image
Parse pull out used AC VLCs
Choose Metadata
Embed watermark

• The three major components of the JPEG
  watermarking tool are
• Parsing the image and pulling out AC VCLs that
  occur in the image
• Using binary tree structure to determine
  watermark bit locations
• Watermark embedding

31
JPEG Results

• Algorithm applied to Lena image
• Image is grayscale, JPEG compressed at quality
  factor of 90
• Image is 45.6 Kb
• Uses the example AC VLC table in the JPEG standard

32
JPEG Results

• Algorithm applied to Lena image
• Virtually no change in file size (change on the
  order of bytes due to zero-padding)
• Lossless
• Applied directly in compressed domain

33
JPEG Visualization

• Loss of synchronization for watermark unaware
  decoders
• Example image only has two watermark bits
  embedded
• Large visual distortion, often can not be
  displayed at all

34
The Next Step Visual Masking

• Current method causes large visual distortions
  for standard decoders (unaware of watermark
  algorithm)
• Goal is to modify algorithm to mask watermark by
  maintaining synchronization with standard
  decoders without sacrificing any previous criteria

35
Visual Masking

• To mask visual impact, modify the run/size of the
  Huffman table for unused VLCs that are mapped to.
• Change run/size of watermarked VLCs to match the
  original run/size
• This is constrained by length of VLCs and number
  of expected appended bits (size)

36
Future Work

• JPEG Related Work
• JPEG visualization
• Incorporate JPEG work into stand alone package
• Maximize capacity for JPEG application
• Add additional security features to JPEG work
• Apply general algorithm to other specific
  compression standards H.263, MPEG-1,2

37
References

• 1. G.C. Langelaar et al. Watermarking Digital
  Image and Video Data IEEE Signal Proc. Magazine,
  Vol. 17, No. 5, Sept. 2000, pp. 20-46.
• 2. L. Chun-Shien, J. Chen, H. Liao, and K. Fan,
  Real-Time MPEG-2 Video Watermarking in the VLC
  Domain, International Conference on Pattern
  Recognition, Vol. 2, pp. 552-555, 2002.
• 3. F. Hartung and B. Girod, Watermarking of
  Uncompressed and Compressed Video, Signal
  Processing (special issue on watermarking), Vol.
  66, No.3, pp.283-302, 1998.
• 4. I. Setyawan et al. Low-bit-rate video
  watermarking using temporally extended
  differential energy watermarking (DEW) algorithm
  SPIE Proceedings on Security and Watermarking of
  Multimedia Contents III, San Jose, USA, January
  22-25, 2001, pp. 73-84.
• 5. Takishima, Wada, and Murakami. Reversible
  Variable Length Codes IEEE Trans. on
  Communications, Vol. 43 No. 2/3/4,
  February/March/April 1995 pp.158-162.
• 6. Alattar, Celik, and Lin. Evaluation of
  Watermarking Low Bit-rate MPEG-4 Bit Streams
  SPIE Proceedings on Security and Watermarking of
  Multimedia Contents V, San Jose, USA, January
  21-24, 2003, pp.440-451.
• 7. Lang, Thiemert, Hauer, Liu, and Petitcolas.
  Authentication of MPEG-4 data risks and
  solutions SPIE Proceedings on Security and
  Watermarking of Multimedia Contents V, San Jose,
  USA, January 21-24, 2003, pp.452-461.
• 8. I. Moccagatta, S. Soudagar, J. Liang, and H.
  Chen, Error-Resilient Coding in JPEG-2000 and
  MPEG-4, IEEE Journal on Selected Areas in
  Communications, vol. 18, no. 6, pp.899-914, June
  2000.