1. Embedding Secrets in Digital Images and Their Compression Codes 嵌入機密訊息於數位影像及其壓縮碼之技術
Advisor: Chin-Chen Chang1, 2
Student: Yi-Pei Hsieh2
1 Dept. of Information Engineering and Computer Science, Feng Chia University
2 Dept. of Computer Science and Information Engineering, National Chung Cheng University

2. Motivation - data hiding (1/4)
Illegal user
Information
Public network
Sender
Receiver
2019/2/16

3. Motivation - data hiding (2/4)
Cryptography
Encryption key
Decryption key
Meaningless & distorted
Plaintext
Plaintext
Ciphertext
Encryption algorithm
Decryption algorithm
Public network
Sender
Receiver
2019/2/16

4. Motivation - data hiding (3/4)
Information
Public network
Illegal user
Sender
Information
Receiver
2019/2/16

5. Motivation - data hiding (4/4)
Compressed codes:
…..
…..
Information
Public network
Sender
2019/2/16
Receiver

6. Outline Part I: embedding secrets into digital images
Image hiding
Part II: embedding secrets into compressed codes
Reversible hiding with high capacity in VQ domain
2019/2/16

7. Part I: embedding secrets in digital images
Image hiding: hiding multiple, relatively-large secret images into a relatively-small cover image

8. Image hiding Cover image Stego image Extracted secret image
● Extracted secret image with Acceptable quality
● Stego-image with high quality
● High hiding capacity
Goals
● Compression method
● Modulus substitution
Techniques
Vector quantization
2019/2/16

9. Vector quantization (VQ)X
Euclidean distance
■ How to generate a representative codebook
■ How to search the closest codeword
 Two-codebook combination
2019/2/16
 Three-phase block matching

10. Two codebook combination
The cover codebook
The difference codebook 1 .
N-2
N-1
(43, 57, …, 40)
Cover image
LBG clustering
(k-means clustering) 1 .
M-2
M-1
(-5, 6, …, 10)
Secret image
Difference image
LBG clustering
(k-means clustering)
2019/2/16

11. Three-phase block matchingj k
index j
index k
Secret image
index j,1
index k,3 10 j k 1
index j,0
index k,1 11 j 1 k 3
2019/2/16

12. Embedding Modulus substitution 40:00101000 Hidden bits: 3(11)
The indices of chosen initial vectors (cover codebook)
Modulus substitution
The difference codebook
LSBs of cover image
q LSBs (q=2)
The compressed indices
40:
Hidden bits: 3(11) 3 2 1 00 11 +3 -1
39:
Parameters
2019/2/16

13. Experimental results (1/3)
Test images
Baboon
Tiffany
Scene
2019/2/16
Lena
Jet
Pepper

14. Experimental results (2/3)
Embed a secret image into a cover image (i.e., Baboon) of the same size
The PSNRs of the extracted secret images
The PSNRs of the stego images
2019/2/16
Hu’s scheme
Wang-Tsai’s scheme
The proposed scheme

15. Experimental results (3/3)
Hide multiple images of large size into a small cover image
The first secret image
The second secret image
cover image
Hu’s scheme
Methods
The stego image
The first extracted secret image
The second extracted secret image
Hu’s scheme
48.86
28.92
28.18
Wang and Tsai’s scheme
37.25
29.55
31.11
The proposed scheme
41.05
33.67
33.81
2019/2/16
Wang-Tsai’s scheme
The proposed scheme

16. Part II: embedding secrets in compressed codes
Reversible data embedding with high embedding capacity

17. Reversible data hiding
Public network
Sender
Original codes
…..
Compressed codes:
…..
…..
2019/2/16
Information
Receiver

18. Reversible data hiding
The similarity property of adjacent areas
Declustering
Put dissimilar codewords into the same cluster
Embedding
Cartesian product
2019/2/16

19. Declustering decluster G-1={C0,C7}
Declustering based on minimum spanning tree
Declustering based on short spanning path
2019/2/16

20. Embedding procedure (1/2)
Neighboring pixel intensities in an image are pretty similar.
Seed Block
Residual Block
Declustered result
Seed indices
Side-match distortion
X = (81, 15, 53, 34, 51,?, ?, ?, 91, ?, ?, ?, 49,?, ?, ?)
Index table
If SMD(X, C4)< SMD(X, C1) and SMD(X, C4)< SMD(X, C6)
Non-exchangeable
Exchangeable
Else
2019/2/16

21. Embedding procedure (2/2)
Non-exchangeable
Exchangeable
Seed indices
Index table
Modified index table
Declustered result
Secret bits:
( )2 = (20)10
101
2019/2/16

22. Experimental results (1/9)
512×512 test images
Baboon
Barbara
Boat
2019/2/16
Lena
Pepper
Tiffany

23. Experimental results (2/9)
Choose different codewords as the roots in the same minimum spanning tree
Declustered result
2019/2/16
Declustered result

24. Experimental results (3/9)
Comparison of selecting different roots as the roots in the minimum-spanning-tree
Baboon
Barbara
Boat
＊ Use the codebook with 1024 codewords ＊Randomly choose codewords C58, C729, C134, C894, and C341 as roots
2019/2/16

25. Experimental results (4/9)
Comparison of selecting different roots as the roots in the minimum-spanning-tree
Lena
Pepper
Tiffany
2019/2/16

26. Experimental results (5/9)
Declustering
The minimum-spanning-tree algorithm
The short-spanning-path algorithm
2019/2/16

27. Experimental results (6/9)
Comparison between using the minimum-spanning-tree and the short-spanning-path algorithms
Baboon
Barbara
Boat
2019/2/16
＊Use the codebook with 1024 codewords

28. Experimental results (7/9)
Comparison between using the minimum-spanning-tree and the short-spanning-path algorithms
Lena
Pepper
Tiffany
2019/2/16

29. Experimental results (8/9)
Comparison of embedding capacities (bits)
Methods
Proposed method
MFCVQ
Chang et al.’s method
Chang and Lin’s method
Lena
36,288
5,892
16,129
8,707
Pepper
36,414
5,712
8,421
Baboon
31,272
1,798
3,400
Barbara
33,222
5,019
5,572
Boat
35,482
6,278
7,838
Tiffany
36,960
6,020
8,830
＊Use the codebook with 512 codewords ＊declustering using the short-spanning-path algorithm (101 groups)
2019/2/16

30. Experimental results (9/9)
Comparison of time required for embedding process
Methods
Proposed method
MFCVQ
Chang et al.’s method
Chang and Lin’s method
Lena
0.23
1.36
352.13
22.80
Pepper
0.26
1.356
351.83
23.1
Baboon
0.21
1.282
352.57
23.2
Barbara
0.22
1.348
352.04
23.6
Boat
0.25
1.357
351.25
22.9
Tiffany
0.28
1.324
352.52
24.0
2019/2/16

31. Conclusions
To hide multiple, relatively-large secret images into a relatively-small cover image
We have proposed an image hiding method with two-codebook combination and three-phase block matching
To develop a reversible hiding with high capacity for VQ domain
We have proposed two declustering methods
We have applied the similarity property of adjacent areas in a natural image and Cartesian product
2019/2/16

32. Future Research Directions (1/2)
Reversible data hiding
Restore the Original cover images after extracting the secret data
Need no extra data
Enhance the embedding capacity
Design the data hiding methods to other image formats
Binary and color images
2019/2/16

33. Future Research Directions (2/2)
Develop the hiding schemes using other compressed codes
JPEG, JPEG2000, and BTC (block truncation coding) compressed codes
2019/2/16

34. Thanks for your attention...