Data Hiding in Image and Video: Part II—Designs and

Download Report

Transcript Data Hiding in Image and Video: Part II—Designs and

Data Hiding in Image and Video:
Part II—Designs and Applications
Min Wu, Heather Yu, and Bede Liu
Outlines
Introduction
Multilevel Data Hiding in Grayscale Image
Multilevel Data Hiding in Video
Conclusion
Introduction
Goal:

apply the solutions in Part I to specific design
problems and present details of embedding data
Multilevel Data Hiding in
Grayscale Image
Introduction
Spectrum Partition
System Design
Experimental Results
Multilevel Data Hiding in
Grayscale Image -- Introduction
Present a two-level data hiding using two types of
embedding mechanisms

Basis: Fig5. in Part I
Basic Assumptions/Conditions:




Grayscale Images
Embedding Domain: 8*8 block DCT coefficients
Using Spectrum Segments for Embedding
Dealing with non-coherent case
Multilevel Data Hiding in
Grayscale Image -- Introduction
Spectrum Partition
Data Model and Formula
Experimental Results
Spectrum Partition-Data Model(1)
Embedding:
where
 the watermark {s1, …, sn } is an n-sample known
sequence,
 b: a bit to be embedded and is equally likely to be “-1”
or “+1”,
 di: noise, i.i.d. Gaussian
Spectrum Partition-Data Model(2)
A few considerations

Bits can be embedded in all bands. In many
cases, bits are embedded in mid-band due to
Low band coefficients generally have higher power
 High band coefficients are vulnerable to attacks


Noise Model can be extended to Normal
Distribution with Various Covariance.
Whitening should be performed in such cases
Spectrum Partition-Data Model(3)
The detector
The mean
Spectrum Partition-Simulation(1)
Subject: 141 Images
Embedding: the Block-DCT spread spectrum
algorithm proposed by Podilchuk-Zeng
Detection: the q-statistic proposed by Zeng-Liu
Three watermarks are used
Pre-processing:


An estimation of the host signal’s power is performed
based on testing images
A set of known signals are added to help locating host
signal from noise
Spectrum Partition-Simulation(2)
Detection: Defined two statistics: q’ and q, with
and without the weighting
Spectrum Partition-Simulation(3)
Experiments:


DCT coefficients are ordered in zig-zag order
Several distortion are introduced while computing q-statistics



JPEG with different quality factors
Low pass filtering
q-statistics are normalized with respect to number of embeddable
coefficients, see Figures


Q is maximum when the embedding starts around 6-11
Q’ is larger than q and it’s monotone
Conclusion:


For high robustness, embed the bit to mid-band coefficients
For high payload, embed the bit to low-band coefficients
Spectrum Partition-Simulation(4)
Spectrum Partition-Simulation(5)
Spectrum Partition-Simulation(6)
System Design
Block Diagram
Two Level Embedding
System Design– Block Diagram(1)
Embedding
System Design– Block Diagram(2)
Detecting
Two Level Embedding(1)
First Level:
Using Odd-Even Embedding in the Low Band
 Quantization Techniques are applied

Two Level Embedding(2)
Second Level:
Using Type I Spread Spectrum Technique
 Antipodal Modulation Is Used



where {vi}: original coefficients
{vi’}: marked coefficients
{b’}: antipodal mapping from b, which is +1 or –1
: watermark strength, adjusted by the just-noticeabledifference (JND) standard
Experimental Results
Multilevel Data Hiding in Video
Embedding Domain
Variable Embedding Rate (VER) Versus
Constant Embedding Rate (CER)
Control Data Versus User Data
Experimental Results
Embedding Domain(1)
Problems Introduced by Consecutive Frames



Add/Drop Some Frames
Switch the Order of Frames
Generate New Frames
Possible Attacks

Collusion Attack
Solution

Adding Redundancy
Embedding Domain(2)
To Avoid Frame-Jitter
Partitioning the Video into Temporal Segments
 Embedding Same Data in Every Frame of a
Segment

Embedding Domain(3)
To Avoid Frame Drop, Reordering, Insertion
Embedding the Same User Data As Well As a
Shorten Version of Segment Index
 The Segment Index Is Part Of the Control Bits

Variable Embedding Rate (VER) vs.
Constant Embedding Rate (CER)
Problem

The Uneven Embedding Capacity Arises Both From
Region to Region within a Frame and From Frame to
Frame
Solution

Combine VER and CER


The Intra-Frame Unevenness Is Handled by CER and
Shuffling
The Inter-Frame Unevenness Is Handled by VER and
Additional Side Information
Number of Bits Embedded in Each Frame
Number of Bits That Can Be Embedded in Each
Frame Changes Greatly
Estimate Number of Bits for Each Frame





Estimate the Achievable Embedding Payload Cˆ
Based on Energy of DCT Coefficients, Number of
Embeddable Coefficients
Set Two Threshold  1 and  2
ˆ   do not embed data
If C
1
If
a number of bits are embedded
1  Cˆ   2
If
Cˆ   2
bits are embedded in higher rate
Estimation of Payload
For Type I Spread Spectrum Embedding,
 The Mean of Detection Statistic Is
E (T )
 Bit Error Probability Is Given by Q ( E (T ))


(max)
Maximum Bit Error Probability Is Given by Pe
A Lower Bound of Mean Detection Statistic Is Defined by
Tth  Q1 ( Pe(max) )

The Detection Statistic When All Embeddable Coefficients
Are Used Is Given By T0

The Payload Is
Control Data Versus User Data(1)
Control Data: Additional Information

Include Frame Sync Index, Number of Bits
Embedded in Each Frame
Embedding Frame Sync
A Short Version of Video Segment Index
 Assume Frame Sync’s Range is 0 to K-1
 The i-th Segment Is Labeled as mod(i, K )

Control Data Versus User Data(2)
User Data: Information
TDM with Shuffling IS Applied
Orthogonal Modulation Is Used to Double the
Number of Embedded Bits

Assume 2B bits Are Embedded
Block Diagram
Experimental Results
Conclusion
Demonstrate How to Apply General
Solutions in Part I to Specific Designs
Made use of
Two types of Embedding
 Modulation and Multiplex Techniques
 Shuffling
 Multilevel Data Hiding
