Video Watermarking Real-time Labeling of MPEG-2 Compressed Video G. C. Langelaar, R. L. Lagendijk, and J. Biemond ITS, ICTG, Delft University of Technology.

Video Watermarking

Real-time Labeling of MPEG-2 Compressed VideoG. C. Langelaar, R. L. Lagendijk, and J. Biemond

ITS, ICTG, Delft University of Technology

Multimedia Security

2

Outline

• Introduction to MPEG video compression standards

• Watermarking MPEG compressed video sequence– Labeling in the bit domain– Labeling in the coefficient domain

• Conclusion

Introduction to MPEG Video Compression Standard

4

MPEG Video Encoding

Zig-ZagQuantDCT

DPCM

RLE

Huffman or Arithmetic

Coding

01001...

I-Frame

RG

B

Y

I Q

To Other Color Space(Optional)

ScanIQ

uantID

CT

Reconstruct &Update

Y

I QForward Frame Buffer

Y

I QBackward Frame Buffer

MotionEstatimation

P B-Frame

Different Image

Motion Vector

5

Discrete Cosine Transform

• DCT

– Convert the time domain signal into frequency domain

88 84 83 84 85 86 83 82 86 82 82 83 82 83 83 81 82 82 84 87 87 87 81 84 81 86 87 89 82 82 84 87 81 84 83 87 85 89 80 81 81 85 85 86 81 89 81 85 82 81 86 83 86 89 81 84 88 88 90 84 85 88 88 81

675 1 -6 2 -2 0 5 -5 -4 1 2 1 5 1 -3 0 2 3 4 6 -2 2 1 5 -3 -1 0 2 0 -2 2 -4 4 3 1 -1 -2 1 -3 1 1 -2 0 -3 2 -1 1 1 3 0 -1 0 -1 -1 0 -2 -1 -1 -5 5 2 -2 2 0

DCT

1)(,2

1)0(

),16

12cos(

16

12cos(),(

2

)(

2

)(),(

7

0

7

0

xcc

lm

kn

mnxlckc

mnDCTn m

6

Quantization

• Quantization

– lossy data compression technique– e.g. 4/2 = 2, 5/2 = 2

675 1 -6 2 -2 0 5 -5 -4 1 2 1 5 1 -3 0 2 3 4 6 -2 2 1 5 -3 -1 0 2 0 -2 2 -4 4 3 1 -1 -2 1 -3 1 1 -2 0 -3 2 -1 1 1 3 0 -1 0 -1 -1 0 -2 -1 -1 -5 5 2 -2 2 0

168 0 -1 0 0 0 1 -1 -1 0 0 0 1 0 0 0 0 0 1 1 0 0 0 1 0 0 0 0 0 0 0 -1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 -1 1 0 0 0 0

Quant

Xq(i,j)=X(i,j)/Q(i,j)

7

ZigZag Scan

• ZigZag Scan– make all zeros to be concatenated– get compression gain

168 0 -1 0 0 0 1 -1 -1 0 0 0 1 0 0 0 0 0 1 1 0 0 0 1 0 0 0 0 0 0 0 -1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 -1 1 0 0 0 0

Zigzag168,0,-1,0,0,0,0,0,0,0,0,1, 0,1,0,0,0,1,1,...,1,0,0,0,-1, 0,0,0,0,0,0,0,0,0,0

8

Run Length Encode

• Run Length Encode– Encode continuous zeros and a nonzero

into a symbol

RLE(1,-1) (8,1) (1,1) (3,1) (0,1) (9,1) (3,-1) (14,1) (4,-1) (0,1) (3,-1) EOB

168,0,-1,0,0,0,0,0,0,0,0,1, 0,1,0,0,0,1,1,...,1,0,0,0,-1, 0,0,0,0,0,0,0,0,0,0

RLE(x0,x1,x2,…) = (zeros,nozero)...

9

DPCM

• Difference Pulse Coded Modulation– using the previous DC to predict the next DC

value

DC=138 DC=140

DC=136 DC=138

DPCM138 => 2 => -4 => 2

XDPCM= Xi - Xi-1

10

Variable Length Codes

• Variable Length Coding

– code representation length is decided by its appearing possibility– short codeword to represent the high frequency event

ABAC

Fix length code

A : 00B : 01C : 10

00010010

Variable length code

A : 1B : 01C : 00

101100

8 bits 6 bits

11

Motion Estimation

88 84 83 84 85 86 83 82 86 82 82 83 82 83 83 81 82 82 84 87 87 87 81 84 81 86 87 89 82 82 84 87 81 84 83 87 85 89 80 81 81 85 85 86 81 89 81 85 82 81 86 83 86 89 81 84 88 88 90 84 85 88 88 81

84 82 83 81 85 86 83 81 82 82 81 83 82 83 83 81 83 82 84 87 87 87 81 88 81 85 86 88 82 82 84 87 81 84 85 87 85 89 84 81 82 85 81 84 81 89 81 83 81 87 86 83 86 89 81 84 88 82 87 84 87 89 84 81

Frame N Frame N+1

-4 -2 0 -3 0 0 0 -1 -4 0 -1 0 0 0 0 0 1 0 0 0 0 0 0 4 0 -1 -1 -1 0 0 0 0 0 0 2 0 0 0 4 0 1 0 -4 -2 0 0 0 -2 -1 6 0 0 0 0 0 0 0 -6 -3 0 2 1 -4 0

- =

Motion Vector

12

Motion Compensation

• Motion Compensation

– using motion vector between current frame and reference frame to reconstruct the prediction of current frame

Reference frame

Current frame

Reconstruct frame

13

P,B-Frames

• The P-picture uses MC to de-correlate dependence between continuous frames

• The B-picture is introduced for increasing the frame rate without increasing too much bitrate

I or P I or PB

Forwardprediction

Backwardprediction

14

GOP Layer

• I,P,B three type of picture to consist a GOP(group of picture)

Temporal : 1 2 3 4 5 6 7 8 9 10 11 12 13

Picture type : I B B P B B P B B P B B P

Coding seq : 1 3 4 2 6 7 5 9 10 8 12 13 11

15

I,P,B Frame

• Difference between picture types

I Frame P Frame B Frame

Compression Ratio

Low Good Best

Random Access

Best Hard Hardest

Complexity normal high highest

Watermarking MPEG Compressed Video Sequence

17

Motivation

• Digital video recording devices is entering market.– The development and success of such

devices depend on technological advances and the existence of adequate copy protection methods

18

Simple One-Copy Only Scenario

Label Presented?Yes

Stop Recording

NoAdd Label

Store

Labeled or Un-labeled Data

Labeled Data

Detection/Labeling Module in Video Recorder

19

Requirements of Video Labeling

• Fidelity– The quality of video should not be significantly affected

• Robustness– Against re-encoding at a different bit-rate

• Blind-detection• Directly embedding into the video data

– For different platforms, interfaces, and data file formats• Complexity

– Real-time embedding/detection– Cost of labeling modules in consumer players

• Bit-stream length increasing is forbidden– To avoid hardware buffer overflow

20

Labeling on Compressed Data

• Decoding, labeling, and re-encoding is computation-consuming

• A labeling scheme for MPEG video should avoid – Requiring a full-decoding operation– Drastically decreasing the quality of video– Increasing the size of labeled video data

• We need techniques which can be applied directly on the MPEG compressed data!!

21

Labeling MPEG Video Streams

• A real-time labeling algorithm for compressed video should closely follow the compression standard to avoid computationally demanding operations– DCT/IDCT– Motion estimation

• The algorithm should work on the lowest layer– Quantized DCT-block

22

DCT-block Representing Domains

Labeling in the Bit Domain

24

Bit Domain Labeling

• To label bit-stream L, consisting of bits Li (i=1,2,…,l) in the MPEG stream – Suitable VLCs are selected – The LSB of their quantized level are forced to

the value of Li

25

Selecting Suitable VLCs

• Label-bit carrying VLC (lc-VLC)– Codewords which another codeword exists wit

h• the same run length• a level difference of 1• the same codeword length

• Such a change of VLC will yield perceptually invisible degradations after decoding

26

Example of lc-VLC

27

Embedding in the Bit Domain

Test codewordsin a macroblock

An lc-VLC codeword is found

If the LSB of its levelequal to the label bit Li?

The codeword is unchanged

Yes

NoReplace the lc-VLC codeword with an

other one

The procedure is repeated until all label bits are embedded

28

Extraction in the Bit Domain

• The procedure is repeated until for i=1,…,l

Test codewordsin a macroblock

An lc-VLC codeword is found

Assign the value represented by its LSB to Li

29

The Label Bit-rate

• The maximum number of label bits that can be added to the video stream per second– Determined by the number of lc-VLC codewor

ds – Not known in advance– Experimental evaluation is required first

30

Evaluating the Label Bit-rate

• A MPEG-2 video sequence is tested – 10 seconds – 720x576 pixels– 25 frames per second– GOP length is 12– Containing I, P, B frames

31

Evaluating the Label Bit-rate (cont.)

Intra coded macro-blocks only

Intra and inter coded macro-blocks only

32

Video Quality Degradation

• Informal subjective tests– The labeling does not introduce any visible

artifacts in 4, 6, and 8 Mb/s coded streams– As for coding at a bit-rate less than 2 Mb/s,

the unlabeled bit-stream already contain too many coding artifacts

33

Video Quality Degradation (cont.)

• Numerically measurements– Difference images are amplified by 60

Unlabeled I-frame Frame difference (4 Mb/s) Frame difference (8 Mb/s)

Most differences are located around the edges and in the textured areas

34

Distributions of the lc-VLC Codewords

• According to the experiments, it appear that the lc-VLC are fairly uniformly distributed over the DCT spectrum– Each non-zero DCT-coefficient represented b

y a VLC have an equal probability of being modified

• Why this conclusion? Smooth areas has fewer nonzero DCT coefficients, and thus fewer VLC codewords

35

Distributions of the lc-VLC Codewords (cont.)

• The maximum local degradation should be as low as possible – The number of lc-VLC codewords per block should be limit

ed

87% of all blocks

64% of all lc-VLC11% of all blocks

36

Distributions of the lc-VLC Codewords (cont.)

• Using a threshold mechanism to limit the number of lc-VLC replacements per DCT block to Tm

The effect of limiting the number of lc-VLC replacements per block

37

Drifts due to Prediction

• ∆MSE=MSEl-MSEu

– MSEl: the MSE per frame between the uncompressed sequence and the labeled sequence coded at 8 Mb/s

– MSEu: the MSE per frame between the uncompressed sequence and the sequence coded at 8 Mb/s

38

Robustness of Bit Domain Embedding

• Bit domain is efficient and simple, but it can also be removed without significantly affecting the quality of video– Decoding and re-encoding using another bit-

rate– Labeling the video stream again using

another random label bit-stream • Suitable for consumer applications

Labeling in the Coefficient Domain

40

Coefficient Domain Labeling

• Coefficient domain labeling embeds the label bit-stream L consisting of bits Li, i=1,…,l, in the I-frames only.

• Each bit out of the label string has its own label-bit carrying region, lc-region, in a I-frame.

41

Example of lc-region

• The first label bit is located in the top-left corner of a I frame

42

Embedding in the Coefficient Domain

• A label bit is embedded in a lc-region by introducing an energy difference between the high-frequency DCT-coefficients of the top-half of the lc-region (denoted by lc-subregion A) and the bottom-half (denoted by B)

• Definition of the total energy in A

• The energy difference D

)}(|}63,0{{)(

)()(12/

0 )(

2),(

cuucS

DCTcoefcEn

b cSubuA

BA EED

S(c) is defined by thecut-off point c

43

Embedding in the Coefficient Domain (cont.)

• Label bit ”0” is defined as D>0, and label bit “1” as D<0

44

Embedding in the Coefficient Domain (cont.)

• If a bit “0” must be embedded, all energy after the cut-off points in the DCT-blocks of lc-subregion B is eliminated by setting the corresponding DCT-coefficeints to zero. Thus,

• If label bit”1” must be embedded, all energy after the cut-off-point in the DCT-blocks of lc-subregion A is elimated, so that:

AABA EEEED 0

BBBA EEEED 0

45

Why Using this Energy Difference?

• The coefficients can be forced to zero by shifting the EOB marker towards the DC-coefficient

46

The Side Effects of Coefficient Elimination

• Since coefficients are removed to add a label, the labeled compressed video stream will always be smaller than the unlabeled video stream.

• If it is necessary to keep the original size of compressed video sequence, stuffing bits can be inserted .

47

Factors Determining EA and EB

• The spatial contents of the sub-regions A and B

• The cut-off point C

48

Determining the Cut-off Points

• Labeling

• Extraction

)})(())((|}63,0{{

),(

DcEDcEcc

cMaxc

tmpBtmpAtmpT

T

)})((|}63,0{{),(

)},)((|}63,0{{),(

),,(

TcEcccMinc

TcEcccMinc

ccMinc

tmpBtmpTTB

tmpAtmpTTA

BA

D and n together determines the visibility and robustness

0<T<D

49

Labeling Procedure

50

Label Extracting Procedure

51

Discarded Coefficients

• D=20, T=15 • Not enough high frequency coefficients exist in the 1.4 Mb/s

compressed stream, since only 75% of the extracted label bits are correct.

52

Video Quality Degradation

Unlabeled I-frame Frame difference (4 Mb/s)Frame difference (8 Mb/s)

• Numerically measurements– Difference images are amplified by 60– Fewer difference per frame, but larger difference per

block

53

Histogram of the cut-off Points

• The lower the bit-rate is, the lower the cut-off points have to be because of the lack of high frequency components in the compressed video stream

54

Discarded VLC Codewords Per Block

95% of all blocks

Coded at 8 Mb/s

10% of blocks have energy above the cut-off point

55

Limiting Discarded VLC Codewords

• Some coefficients will not be eliminated due to the limit

• Error-correction codes can be used

56

MSE of different Frame Types

The sequence is coded at 8 Mb/s The degradation now fades out for P and B frames

57

Robustness of Coefficient Domain Labeling

• Possible attacks– Re-encoding– Trans-coding

• Trans-coding a 8Mb/s sequence at a lower bit-rate

58

Conclusions

• Comparison of the two labeling schemes– The bit domain labeling method is

computationally highly efficient and a high label bit-rate, but not robust against re-labeling or trans-coding

– The coefficient domain labeling is more complex and has a lower bit-rate, but has higher robustness

59

Conclusions (cont.)

• Both schemes are suitable for a real-time copy protection system for digital video recording devices

60

The European ACTS SMASH Project

• The problem which labeling method should be used for a digital video recording device is currently under investigation within the European ACTS SMASH project

61

References for video watermarking

• 1. Gerrit C. Langelaar and Reginald L. Lagendijk, “Optimal Differential Energy Watermarking of DCT Encoded Images and Video,” IEEE Trans. On Image Processing, vol.10, No.1, Jan. 2001, pp.148-158

• 2. Real-Time Labeling of MPEG-2 Compressed Video – http://ict.ewi.tudelft.nl/pub/inald/realtime.pdf