1. A Robust and Recoverable Tamper Proofing Technique for Image Authentication
Authors: Chin-Chen Chang & Kuo-Lung Hung
Speaker : Chin-Chen Chang

2. Problem Definition The definition of tampered detection and recovery
When somebody modified a specific image, the technique can detect the tampered areas, and, moreover, can recover the modification.

3. Motivation
Original Boat Image
Tampered Boat Image

4. Motivation
Boat Image after Detection
Boat Image after Recovery

5. Requirements Effectiveness: Differentiation: Security: Recoverability:
Provide a high probability of tamper detection
Can distinguish between an innocent adjustment and replacing or adding features on purpose
Only a selected group of people having a secret key can perform the detection
Can recover back to the correct image from the modification

6. The Proposed Method Three techniques are employed Watermark Generating
Watermark Embedding
Tamper Detecting and Recovering

7. Watermark Generating Divide the image into 88 blocks. 8 8

8. Watermark Generating (cont.)
Divide a block into 44 sub-blocks. 8 8

9. Watermark Generating (cont.)
Calculate the mean values of the sub-blocks of a block to form a 4-dimensional vector. 8 8 .
101 89 59 67 8 8 88
108 78 97
198
159 74 24 4
169
124 57 33 . . 4

10. Watermark Generating (cont.)
50,63,50,142 1 2 3 .
511
Code book
47,50,61,53
127,200,37,83
Generate a local codebook
5. Encode each block
Index table 23 29 25 34 12
475 46
243
325 73
148
369
257 1 56 48
. . .
Watermark = (the index table + checksums of the index table)
+the RS-encoded local codebook
16 bit/per block

11. Watermark Embedding Scramble the watermark
DCT coefficients
Scramble the watermark
Hide the watermark to the middle-frequency DCT coefficients of each block

12. Watermark Embedding (cont.)
Bit hiding
Qa and Qb in Q are two values corresponding to ma and mb in DCT transformed block
Adjust (ma,mb) to (m’a,m’b)
if ej=1
if ej=0
,where and , q is the
compression factor of JPEG
one solution

13. Watermark Embedding (cont.)
Bit hiding (example)
{m}= (10, 15, 22, 38)={m1,m2,m3,m4}
{w} = {e1,e2}={1,0}
(Q1,Q2,Q3,Q4) = (14, 16, 13, 14) q=10 => (1, 2, 3, 4) = (2,3,2,2)
m1'=((m1/2+m2 /3)/2+1)*2 = 14
m2'= ((m1/2+m2 /3)/2+0)*3 =12
w1=1  m1/1  m2 /2
m3'= ((m3/2+m4 /2)/2+0)*2 = 30
m4'= ((m3/2+m4 /2)/2+1)*2 = 32
w2=0  m3/3 < m4 /4

14. Detection and Recovery
Extract the hidden watermark
Use the error correction technique to correct the watermark
Obtain the hidden mean value M’i of each sub-block from the extracted watermark
Calculate Abs(Mi-M’i )
Make judgement by
tampered if Abs(Mi – M’i) > threshold,
not tampered otherwise.

15. Error Correction Assumption Correct the codewords Correct the indices
Only 1 bit error in an index or a codeword (the probability 96%)
Correct the codewords
Use the RS coding to correct one bit error.
Correct the indices
Check the checksum: if not correct, the following correction is executed.
Generate the set of 1-Hamming distance candidate indices which include the index itself.
For each candidate index, obtain the codeword Ci.
Calculate four mean values of the block to form a vector X.
Choose the one who has the minimum distortion between Ci and X.

16. Experimental Results SRTPT Our method =4 =6 =8 q=5 q=10 q=15
PSNR(dB) of the embedded image
36.791
33.109
30.398
37.038
33.256
30.762
% of error blocks at 20% contrast
25.17%
17.16%
15.33%
11.00%
11.10%
11.08%
Recovery PSNR at 20% contrast
21.893 24
24.467
27.006
26.289
25.441
% of error blocks at one blur
21.06%
17.79%
16.07%
12.08%
1.23%
0.08%
Recovery PSNR at one blur
18.589
19.909
19.621
20.501
29.666
33.605

17. % of error blocks at 3% noise 43.41% 13.54% 2.84% 38.22% 4.46% 0.52%
SRTPT
Our method
=4
=6
=8
q=5
q=10
q=15
% of error blocks at 3% noise
43.41%
13.54%
2.84%
38.22%
4.46%
0.52%
Recovery PSNR at 3 % noise
15.507
23.561
29.675
16.024
24.326
29.367
% of error blocks of JPEG (1:3)
7.31%
0.35%
0.27%
0.92%
0.00%
Recovery PSNR of JPEG (1:3)
25.429
31.548
30.582
27.904
32.881
31.328
% of error blocks of JPEG (1:4)
42.23%
23.00%
10.88%
66.61%
0.05%
Recovery PSNR of JPEG (1:4)
16.884
20.051
25.342
12.858
31.68
32.064
% of error blocks of JPEG (1:5)
52.17%
40.93%
40.08%
78.23%
66.16%
Recovery PSNR of JPEG (1:5)
16.483
16.782
16.966
11.167
12.748
26.376

18. Experimental Results (1)
(a) Embedded image Lena
(b) Contrasted image

19. Experimental Results (1)
(c) Detected image
(d) Recovered image

20. Experimental Results (2)
(a) Original Boat Image
(b) Tampered Boat Image

21. Experimental Results (2)
(c) Boat Image after Detection
(d) Boat Image after Recovery

22. Conclusions
Can effectively detect and recover the image that tampered with
Satisfy the differentiation and robustness requirements