1. A New Method for Tamper Detection and Recovery
Speaker : Kuo-Lung Hung (洪國龍)
Date :

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

3. Requirement Effectiveness: Differentiation: Security: Recoverability:
Provide a very high probability of tamper detection.
Can distinguish between an innocent adjustment and replacing or adding features.
Provide the security mechanism; only a selected group of people sharing a secret key should perform the detection.
Can recover back to the correct image from the modification.

4. Previous Work Problem The security is not enough.
Using the indices of the codewords of size 16(for detection) and the low-bit rate compressed image, namely the indices of the codewords of size 256 (for recovery) as the watermark.
To hide the watermark to the middle-frequency DCT coefficient.
The detection codewords of a block are hiding in the itself block, and its recovery codewords are hiding in the randomly selected block.
Problem
The security is not enough.
The number of hiding bits is too many (= 32 bits/per block). Therefore, the method can not endure some image adjustment operations.

5. The proposed method Three techniques employed by our method:
Watermark Generating
Watermark Embedding
Tamper Detecting and Recovering

6. Watermark Generating 8
Divide the image into 88 blocks.
Divide a block into 44 sub-blocks. 4 45 48 50 47 53 52 55 51 46 80 78 60 65 85 90
Calculate the mean values of the sub-blocks of a block to form a 4-dimensional vector.

7. Watermark Generating (cont.)
Use these vectors to generate local codebook of size 512.
40,50,50,48
50,50,50,55 1 2 3 .
511
Perform the pseudo-gray coding upon the codebook.
50,50,50,55 1 2 3
40,50,50,48 .
511
After pseudo-gray coding
The codebook is used to encode the vector of each block, and some indices are obtained.
Use the RS coding to encode the codewords of the codebook. (for error correction)
Watermark = the indices + checksums of the indices
+the RS-encoded codewords
 16 bit/per block

8. Image Adjustment Before Embedding
The distortion exists between the mean value of each sub-block and the value after VQ decoding. 50 40 45 35
the value after VQ decoding : 48
Motivation
Image adjustment
Calculate the mean value of each sub-block.
the mean value : 45
Subtract the mean value and add the value after VQ decoding to each pixel. 53 43 48 38

9. Watermark Embedding DCT coefficients 1 2 6 7 15 16 28 29
Randomly hide the scrambled watermark to the middle-frequency DCT coefficients of each block.
DCT coefficients
Totally scramble the watermark.

10. Watermark Embedding (cont.)
=2
Example:
Bit hiding mj
H(mj,wj) = 4
4+2  if wj=1,
Mj+2 
4 if wj=0,
H(mj,1) = (mj / 8)  8 + 4
H(mj,0) = ((mj +4)/ 8)  8
eg. H(0,1)->4, H(6,1)-> 4, H(10,1)->12
where   1 is the adjusting magnitude.
eg. H(0,0)->0, H(6,0)-> 8, H(10,0)->8
Bit extracting
E(mj) =
1 otherwise.
0 if ((mj+) mod 4 ) < 2 ,

11. Detection and Recovery
Use the bit extracting function to extract the hided watermark.
Since the extracted watermark might have some errors, the error correction technique needs performed upon the watermark.
According to the extracted watermark, the hided mean value Mi of each sub-block could be obtained.
Calculate the mean value Mj of the sub-block that might have been tampered with. Therefore,
tampered if Abs(Mi - Mj) > threshold,
not tampered otherwise.
the block is:

12. The Demo

13. Remarks
The proposed tamper proofing technique can securely and effectively detect and recover the image that tampered with.
Compare to previous work, the security is enhanced, moreover, the proposed technique can distinguish between an innocent adjustment and intentional modification.
The image adjustment may loss some image quality.
If the range of the modification is large, the detection has some false drops.
To alleviate the problem of (3) and (4) is the future work.