Yun CAO Xianfeng ZHAO Dengguo FENG Rennong SHENG Video Steganography with Perturbed Motion Estimation.

Yun CAOXianfeng ZHAODengguo FENG

Rennong SHENG

Video Steganography with Perturbed Motion Estimation

Outline

Performance

Perturbed Motion Estimation

Motivation

Introduction

Video Steganography

• Adequate payloads

• Multiple applications

• Advanced technologies

Video Steganography

Conventional methodsDomain utilized --Intra frame

--Spatial domain (pixels)

--Transformed domain (DCT)

Disadvantages --Derived from image schemes

--Vulnerable to certain existing steganalysis

Video Steganography

Joint Compression-EmbeddingUsing motion informationAdopting adaptive selection rules --Amplitude

--Prediction errors

Motivation

Arbitrary Modification

Degradation in Steganographic

Security

Known/Week Selection rule

Motivation

How to improve?Using side information --Information reduction process

--Only known to the encoder

--Leveraging wet paper code

Mitigate the embedding effects --Design pointed selection rules

--Merge motion estimation & embedding

Typical Inter-frame Coding

01011100…

Entropy Coding

DCT & QUANTIZATION

Inter-MB Coding

MB PARTITION

Regular Motion Estimation

MB COORDINATE

RC

12,8 4,4

MOTION VECTOR

8,4v

OthersCSimilarityRCSimilarity ,,

Perturbed Motion Estimation

MB COORDINATE

RR’C

12,8 14,7

MOTION VECTOR

8,4v

4,4

10,3'v

',, RCSimilarityRCSimilarity

1' vPvP

C is applicable

Capacity

Number of applicable MBsFree to choose criteriaSAD, MSE, Coding efficiency, etc

Wet Paper Code

Applicable MBs (Dry Spot)

Confine modification to them using wet paper code

Embedding Procedure

Determine Applicable MBs

Wet Paper Coding

Perturb Motion Estimation

Video Demo

Sequence:“WALK.cif”Duration: 14 sMessage Embedded: 2.33KBPSNR Degradation: 0.63dB

Experimental Date

20 CIF standard test sequence352×288 ， 396 MBsEmbedding strength: 50

bit/frame

Preliminary Security Evaluation

Traditional SteganalysisA 39-d feature vector formed by

statistical moments of wavelet characteristic functions (Xuan05)

A 686-d feature vector derived from the second-order subtractive pixel adjacency (Pevny10)

SVM with the polynomial kernel

Preliminary Security Evaluation

Xuan’s Pevny’s

TN TP AR TN TP AR

59.7 39.2 49.5 48.3 53.5 50.9

Preliminary Security Evaluation

Motion vector mapVertical and horizontal components as

two imagesA 39-d feature vector formed by

statistical moments of wavelet characteristic functions (Xuan05)

SVM with the polynomial kernel

Preliminary Security Evaluation

Horizontal Component Vertical Component

TN TP AR TN TP AR

91.5 10.8 51.2 53.5 46.9 50.2

0 0.2 0.4 0.6 0.8 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False Positives

True

Pos

itive

s

0 0.2 0.4 0.6 0.8 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False PositivesTr

ue P

ositi

ves

Preliminary Security Evaluation

Target SteganalysisA 12-d feature vector derived from the

changes in MV statistical characteristics (Zhang08)

SVM with the polynomial kernel

Preliminary Security Evaluation

Zhang’s

TN TP AR

50.5 51.8 51.2

Summary

• Joint Compression-Embedding

• Using side information

• Improved security

Future works

Minimize embedding impactsDifferent parity functionsDifferent selection rule designing criteria

Further SteganalysisLarger and more diversified database