afosr program review data hiding in compresed digital video bijan mobasseri, pi dom cinalli, aaron...

AFOSR PROGRAM REVIEW

DATA HIDING IN COMPRESED DIGITAL VIDEO

Bijan Mobasseri, PIDom Cinalli, Aaron Evans, Dan Cross, Sathya Akunuru

ECE DepartmentVillanova UniversityVillanova, PA 19085

June 6-8, 2002Burlington, VT

2002 AFOSR Program Review 2

Outline

• Data hiding/watermarking requirements• Data hiding in compressed video • Using variable length codes for data hiding• Lossless watermarking using resilient-coding• Video authentication through self-

watermarking• Metadata embedding• Open Issues


Background

• This effort is funded by AFOSR to develop algorithms for the creation of smart digital videos

• The project is monitored by AFRL/IFEC• Applications include

• Watermarking for tamper detection, recovery

• Data hiding for covert communications

• Metadata embedding

• Security and access control


Data hiding requirements

• Data hiding must at least meet the following three conditions:

• Transparency

• Robustness or fragility

• Security• Places to hide data are:

• Spatial- pixel amplitudes, LSB, QIM

• Transform domain- spread spectrum, Fourier/wavelet, LPM

• Joint- time/frequency distribution


State of video watermarking

• Video watermarking is strongly influenced by still image watermarking algorithms where video is modeled as a sequence of stills

• Examples include LSB watermarking of raw frames, spread spectrum and 3D-DFT

• Increasingly, however, the native state of video is in compressed format and does not yield itself to simple still frame modeling


The medium

• Understanding the medium is a prerequisite to watermarking it

• Uncompressed NTSC video runs at 168 Mb/sec. MPEG-2 runs at <10 Mb/sec.; a 96% reduction

• Redundancy is at the heart of data hiding. Compressed video leaves precious little space to hide data while maintaining robustness, security and imperceptibility


Distinction with a difference

• We recognize a difference between

• watermarking of compressed video vs.

• compressed video watermarking • The former refers to watermarking of video

which may later be compressed• The later refers to watermarking that is done

entirely post-compression.


MPEG bitstream syntax

DATA HIDING IN VLCs


Label-carrying VLCs

• Variable length codes are the lynchpin of MPEG

• There is a subset of MPEG VLC codes that represent identical runs but differ in level by just one

From: Langelaar et al, IEEE SP Magazine

September 2000


Data hiding in lc-VLC

• The algorithm proposed by Langelaar embeds watermark bits in the LSB of the level of the lc-VLCs


Data hiding capacities:data

Total lc-VLCs Total VLCs24008 204493

655548 4321760

61298 461027

3614 61985

345315 3290373

24412 204362

run 0 0 0 0 0 0 0level 5 6 8 9 10 11 12406974ai1.mpeg 3734 2630 1835 1423 1122 827 1036% occurrence 15.6 11.0 7.6 5.9 4.7 3.4 4.3status.mpeg 136422 90706 60594 46844 36066 32513 25822% occurrence 20.8 13.8 9.2 7.1 5.5 5.0 3.9avalon.mpeg 13080 9506 5420 4353 3405 2865 2341% occurrence 21.3 15.5 8.8 7.1 5.6 4.7 3.8gtscrush.mpeg 1056 700 281 246 193 131 143% occurrence 29.2 19.4 7.8 6.8 5.3 3.6 4.0final-days.mpg 78900 54245 30487 24116 18751 15464 12791% occurrence 22.8 15.7 8.8 7.0 5.4 4.5 3.7simp.mpg 5997 4007 2438 1441 1627 1225 807% occurrence 24.6 16.4 10.0 5.9 6.7 5.0 3.3

average % occurrence 22.4 15.3 8.7 6.6 5.5 4.4 3.8

Lossless video watermarking using error-resilient VLCs**B. Mobasseri, “Watermarking of Compressed Multimedia using Error-Resilient VLCs,” MMSP02- in review


The idea:watermark as intentional bit errors

• There has been notable cross currents of late between watermarking and channel coding

• A close look reveals that watermarking of VLCs is essentially equivalent to channel errors.

• Bit errors and watermark bits have identical impact. They both cause bit errors in affected VLCs.

• The difference is that channel errors occur randomly whereas watermark bits can be planted at will and at locations that facilitate detection.


The solution-lossless watermarking

• Embed watermark bits in the VLCs as controlled bit errors

• MPEG-2 VLCs, however, have no inherent error protection. Any bit error will cause detection failure up to the next resynchronization marker

• Bidirectionally decodable codewords are capable of isolating and reversing channel errors

• An interesting side effect of the above hypothesis is that if error-resilient VLCs are successful in reversing bit errors, the outcome would be mathematically lossless watermarking


Two-way decodable VLCs

• MPEG-4 uses RVLCs but Girod(1999) has proposed an elegant design whereby conventional VLCs are made to exhibit resynchronizing property

• To construct resynchronizing VLCs from ordinary VLCs, we first define a packet consisting of N consecutive VLCs

€

P = vlc1, vlc2 ,...,vlcN{ }

where

vlci = bi1,bi

2 ,...,bil i( )

{ },b∈ 0,1{ }

€

C = vlc1, vlc2 ,...,vlcN | 00,...,0[ ]⊕

00,...,0 | vlc1' ,vlc2

' ,...,vlcN'

[ ]

L ≥ max l k( ){ }k

vlc’=fliplr(vlc)


Code structure

• Each VLC is represented twice in the new bitstream. It is this property that allows error resiliency

• Burst error shall not be so long to simultaneously affect the same bit of identical VLC


Watermarking using bidirectional codes

€

vlc _ a = a1, a2{ },vlc _ b = b1,b2,b3{ },

vlc _ c = c1,c2{ },vlc _ d = d1,d2 ,d3, d4{ }VLCs: Message:{a,b,d,c}

€

C =

a1 ,a2,b1,b2,b3 ,d1 ,d2⊕ a2 ,d3⊕ a1,

d4 ⊕ b3 ,c1 ⊕ b2 ,c2 ⊕b1 ,d4 , d3 ,d2 ,d1,

c2 ,c1

⎧

⎨ ⎪

⎩ ⎪

⎫

⎬ ⎪

⎭ ⎪

L = zeros(1, 6)

€

Cw =

a1, a2 ,b1 ,b2 ,b3, d1⊕w1,d2 ⊕ a2 ⊕w2,

d3⊕ a1⊕w3, d4 ⊕ b3,c1⊕ b2,c 2⊕ b1,

d4 , d3 ,d2 ,d1 ,c2,c1

⎧

⎨ ⎪

⎩ ⎪

⎫

⎬ ⎪

⎭ ⎪

Bidirectional VLC

Watermarked w={w1,w2,w3,w4) bidirectional VLC


Watermark detection

• On forward decoding, vlc_a and vlc_b will be correctly decoded. Failure will occur at vlc_d

• On forward direction, correctly decoded symbols are {a,b}. On reverse decoding, correctly decoded symbols are {c,d}.

• The last symbol correctly decoded on the reverse path is the same symbol that failed detection on forward decoding. The

correct symbols are then {a,b,d,c}

€

Cw =

a1, a2 ,b1 ,b2 ,b3, d1⊕w1,d2 ⊕ a2 ⊕w2,

d3⊕ a1⊕w3, d4 ⊕ b3,c1⊕ b2,c 2⊕ b1,

d4 , d3 ,d2 ,d1 ,c2,c1

⎧

⎨ ⎪

⎩ ⎪

⎫

⎬ ⎪

⎭ ⎪


Distance properties

• Each VLC in the C stream appears twice. Therefore, the ith bit of a VLC is separated from its copy by bits given by

• If the watermark burst begins with the last bit(LSB) of the VLC, the burst cannot last longer than min bits.

€

=L + l − (2i − 1)

€

min = L − l + 1, for i = 1

δ max = L + l − 1, for i = l


Watermarking capacity

• Watermarking capacity of a VLC falls under two categories

• L=l, in this case

C=L bits/packet

• L>l, watermark burst may cross over to the L-l bits of the next VLC. It follows that

€

capacity = 2L − l


Implementation

Video VLC Statistics# of

frames# of

slicesAvg# of slices/fr

Avg# of VLC/slice

Capacity(bits)

Video#1

871 7839 9 178 133,263

Video#2

8466 8466 1 550 143,922

Video#3

396 396 1 3051 6,732

SELF-WATERMARKING*

*D. Cross, B. Mobasseri, “Watermarking for self-authentication of compressed video,” IEEE ICIP2002, September 22-25, 2002,

Rochester, NY.


Self-watermarking:the concept

• In self-watermarking, the watermark is extracted from the source itself

• Self-watermarking prevents watermark pirating and may allow recovery of tampered material such as cut and paste or re-indexing attacks

• Most work on self-watermarking has been done on images. If it has been done video, the approach is to model video as a sequence of stills


Self-watermarking of compressed video

1 0

VLC(0,5)

VLC(0,16)

VLC(1,15)

VLC(0,6)

VLC(1,10)

VLC(1,11)

VLC(0,12)


Watermark extraction

• Watermark is extracted from the I frame by zigzag scanning of I frame VLCs and storing in array w

• The number of bits in w must be less than or equal to the number of lc-VLCs in gop. In addition, w must contain integer number of VLCs

€

g = vlc1, vlc2 ,...,vlcm{ }

vlci = b1,b2 ,...,bni{ },bni

∈ 0,1{ }

length(vlci) = ni

w = vlc1' ,vlc 2

' ,...,vlc p'

{ }

vlci' = b1

' ,b2' ,...,bqi

'{ }


Watermark embedding

• To be able to fully embed the I frame into the GOP the following must hold

• Once the mask is generated, the embedding method is as follows

€

qii=1

p

∑ ≤ m

€

wk = perm(I ,key)

LSB(vlci) =LSB (vlci),if LSB (vlci) = wk i( )

wk i( ),otherwise

⎧ ⎨ ⎪

⎩ ⎪


DataStream File Size Frame Size

(MB)

Frame Types

MPEG1 892kB 24x16 13I,78B,65P

MPEG2 2257kB 45x36 3I,8P,20B

Frame Size GOP 1 LCVLC #

GOP 2 LCVLC #

GOP 3 LCVLC #

372x246 2997 3008 3061

720x576 32680 11279 14463


Stream Average Capacity (bits/frame)

MPEG1 I = 2958, P = 13, B = 0

MPEG2 I = 18125, P = 501, B = 2

Watermarking capacity

• I frames hold almost all of the watermark data. These results are expected since only the intra-coded macroblocks will hold watermark data.

Metadata Embedding


Background

• Video images & metadata recorded and handled as two separate streams

• Storage overhead

• Bookkeeping issues

• Accuracy and human error

• Cumbersome to display

• It would be nice to permanently attach metadata to video and make it available during playback

MetadataVideo


Metadata WatermarkingVideoBufferVideoBuffer

WMetadataBuffer

MetadataBuffer

MPEGEncoder

Watermarked Video

Store

Display

Watermarking system combines bothvideo and metadata feeds to form a

single, less cumbersome stream thatcan be both displayed and stored.


Implementations• Real-Time Processing

• Metadata is embedded into MPEG video during the recording process and is available for immediate transmission from UAV.

• Batch Processing

• Video & metadata recorded in their entirety before embedding process of metadata into video begins. Data cannot be displayed until watermark process has completed.


Sample Metadata and video footage

Surveillance VideoXML Coded Metadata


Display Utility

• JAVA based application that simplifies display of video & metadata

• Abstracts user from separation of video & metadata


Open Issues

• Open problems in RVLC watermarking are

• Capacity

• Security

• Channel bit errors

• Non-burst errors

• Forced invalidity


T H E E N D

afosr program review data hiding in compresed digital video bijan mobasseri, pi dom cinalli, aaron...

Documents

requirementsdata hiding

image watermarking algorithms

lsb watermarking of

ituncompressed ntsc

native state of video

compressed format

vlcsvariable length

identical runs