1. Robust Image-Adaptive Data Hiding
Kaushal Solanki
Vision Research Lab
Dept. of Electrical and Computer Engineering
University of California, Santa Barbara

2. Today’s talk High Volume Data Hiding (Part I)
Quantization index modulation for images
Adapting to control perceptual degradation
Containing fallout from adaptation using channel coding
Tamper Detection and Localization
Gracefully improving Image-in-image hiding (Part II)
Joint source-channel coding for hiding image in image
‘Print and Scan’ Resilient Data Hiding (Part III)

3. High Volume Data Hiding
PART I
High Volume Data Hiding

4. Methodology Image Adaptive Criterion
Divide image into 8x8 non overlapping blocks
Divide by JPEG quantization matrix
2D DCT
DCT “Coefficients”
Choose coefficients to hide
Image Adaptive
Criterion
Image
Hide using choice of scalar quantizer
Quantize to odd values to hide ‘1’
Quantize to even values to hide ‘0’
Scaling and Inverse DCT

5. Image-adaptive hiding
“Well-known” principles
Hide in transform domain (we use DCT coeffs.)
Hide in low and mid frequencies (robust)
Scalar quantization based hiding ``good enough’’ (2 dB off in Costa setting)
Goal: survive JPEG at design quality factor
Use quantizers specified by JPEG at design quality factor
New insight
Local adaptation: avoid hiding in small coefficients
Other possible constraints
May want to avoid hiding in sensitive parts of image

6. Local vs. Statistical Criterion
Local Criterion
16,384 bits hidden
16,172 bits hidden

7. Basic hiding method Decide on frequency band in which to hide
If coefficient in band larger than a threshold, use quantization index modulation to hide
Also could skip coeffs in sensitive areas
Decoder must guess where encoder has hidden data
Wrong guesses can cause catastrophic sync errors

8. The insertion/deletion problem
Coefficient values change under attack
Insertions: Decoder assuming data has been embedded where there is no data.
Deletions: Decoder assuming data is not embedded where there is data.
Any image adaptive data hiding scheme is vulnerable to this problem

9. Erasures at the encoder
Powerful erasures and errors correcting code
Codeword spans entire set of candidate embedding coefficients
Coefficients where encoder decides not to embed treated as erasures at the encoder
Low-rate RA code works great
Decoder uses same criteria as encoder to guess hiding locations
Insertions become errors
Deletions become additional erasures

10. Decoding
Hard-decision decoding used in general (no need to know attack statistics)
JPEG, wavelet compression, image resizing & tampering
Soft decision decoding used to compare against info-theoretic limits for AWGN attacks

11. Zero-threshold SEC scheme example
Original Peppers Image
Hidden Image (11,073 bits), DQF=25
All bits recovered successfully after ~ 0.42 bpp JPEG compression

12. Zero-threshold SEC scheme example (II)
Original Crowd Image
Hidden Image (30,710 bits), DQF=50
All bits recovered successfully after ~ 0.6 bpp JPEG compression

13. Unity-threshold SEC scheme example
Original Harbor Image
Hidden Image (7,843 bits), DQF=25
The hidden image is virtually indistinguishable from the original!

14. ‘2’ - threshold SEC scheme example
Original Harbor Image
Hidden Image (3,816 bits), DQF=25
Again, the hidden image is virtually indistinguishable from the original!

15. Wavelet compression attack
JPEG 2000 compression was used to attack images with data hidden using the RA coded SEC scheme. The results shown below is for a 512x512 Lena image (design QF=25).
Attack Compression
Number of bits hidden
RA code rate
(1/q)
0.800
7,446
1/11
0.530
4,096
1/20
0.400
2,730
1/30

16. Image Tampering Examples
20 % of 512x512 Lena image tampered. All the 5,820 hidden bits recovered.
Lena image tampered globally All the 6,912 hidden bits recovered.

17. 6,301 bits hidden against image tampering
Tamper Localization
6,301 bits hidden against image tampering
After decoding is performed, the tampered region can be automatically localized

18. Image Resizing
Performance of RA coded SEC scheme for 512x512 Lena image under image resizing attack using bicubic interpolation:
Fewer bits can be hidden against other interpolation methods.
Percentage Resizing
Number of bits hidden
RA code rate
(1/q)
10 %
7,446
1/11
15 %
6,826
1/12
20 %
6,301
1/13

19. Open Issues Detectability Further robustification
So far we have optimized against attacks, not to elude detection
Detection theory
Practice
Further robustification
Geometric attacks

20. Gracefully Improving Image-in-Image Hiding
PART II
Gracefully Improving Image-in-Image Hiding

21. Image-in-image Hiding
Graceful Improvement
Quality of recovered hidden image should be better if attack is less severe
Approach
Hybrid Digital-Analog Hiding Scheme
Analog Information Hiding
MMSE Decoding for JPEG Attacks

22. Graceful Improvement
Quality of the recovered signature image should be better if the attack is milder.
Motivation:
Attack level seldom known a priori
Broadcast scenario: Multiple receivers with different attack channels
Method: Joint source-channel coding
Ideas similar to those on AWGN channel

23. Hybrid Digital-Analog Hiding
Signature image divided into digital data and analog residue.
Source coder
Signature Image - +
Digital Data
Analog Data
RA coding
Hide using SEC scheme
Hide analog information by replacing the residue

24. Hiding Analog Information
To hide an analog number m into a host sample h:
Quantize the host h using a quantizer of step size D
Scale the source m to lie in the interval (0,D)
Replace the residue with the scaled source
Message m always measured from an even reconstruction point.

25. Hiding Analog Information: Example
0.65 1 h
2.25 1 2 3 4 z
2.65

26. Hiding Analog Information: Example
0.65 m 1 h
1.85 1 2 4 3 z
1.35

27. JPEG attacks and MMSE decoding
Varying levels of JPEG compression
Decoder knows attack level (encoder does not)
Minimum mean squared error (MMSE) decoder under uniform quantization attack

28. Image-in-Image Hiding: Implementation
Source coder
Signature Image - +
Digital Data
Analog Data
RA coding
Hide using SEC scheme
Hide analog information by replacing the residue
Processing the signature image
Allocating the channels
Hiding the digital part
Hiding the analog part

29. Processing the Signature Image
Source coder
Signature Image
Digital Data - +
Divide image into 8x8 non overlapping blocks
Analog Data
Divide by JPEG quantization matrix
Huffman Entropy coding D
2D DCT
DCT “Coefficients”
Quantize -
Pre-selected coefficients + A

30. Allocating the Channels
DC Coefficient: Not used for embedding
Band for hiding Analog Information
Candidate embedding band for Digital data
Host Coefficient Block

31. Example 1
Hiding 128x128 image into a 512x512 image with design quality factor (QF) of 25.
Processing the signature image:
Image compressed at QF=10, forming the digital part.
Residues of 16 low frequency coefficients form the analog part.
Allocating the channels:
One coefficient each from each 8x8 block forms the analog channel.
34 coefficients form the candidate band for the digital channel.

32. Example 1 Harbor image with 128x128 peppers image hidden
Original 512x512 Harbor image

33. Received Signature Image: Attack QF = 25 (93.5% compr.) MSE = 0.0286

34. Attack QF = 50 (88.0% compr.) MSE = 0.0128
Received Signature Image:
Attack QF = 50 (88.0% compr.) MSE =

35. Received Signature Image: Attack QF = 75 (81.9% compr.) MSE = 0.0060

36. Received Signature Image: No Attack

37. Example 2
Here, we hide a larger image (a 256x256 image) with a higher design QF of 50.
Processing the signature image:
The signature image is JPEG compressed at QF=12, and residues of 12 low frequency coefficients constitute the analog part.
Allocating the channels
3 coefficients per block are used for sending analog residue and 32 coefficients per block form the candidate embedding band for the digital data.

38. Example 2 Bridge image with 256x256 Lena image hidden
Original 512x512 Bridge image

39. Received Image: Attack QF = 50 (84.3% compr.) MSE = 0.0267

40. Received Image: Attack QF = 75 (76.1% compr.) MSE = 0.0162

41. Received Image: No Attack

42. Open issues
Fundamental limits on joint source-channel hiding in specific contexts
Achieving fundamental limits

43. ‘Print and Scan’ Resilient Data Hiding
Part III
‘Print and Scan’ Resilient
Data Hiding

44. Goal: Surviving Print-Scan
Embed significant volume of information
Say, several hundred bits
Use readily available devices
Ones “sitting” in your lab
Laser printers, flatbed scanners, recycled printer papers
Blind decoding
Preferably, survive other attacks as well
Filtering, resizing, rows/columns removal, etc.

45. Why print-scan?
Interesting and Challenging Problem…

46. Why print-scan, really? Many potential applications…
Document authentication
Security concerns at “all-time high”.
Embed information into pictures in passports, driving licenses, ID cards etc.
Only specific devices with access to a secret key can authenticate.
Forgery becomes very difficult since the hidden information is inseparable from the picture

47. Why print-scan? (Cont.) Potential applications…
Copyright Protection of Images
Many pictures appear in the print media
Can be easily scanned and claimed by others
E-Commerce of Digital Images
High potential, but print-scan resilience necessary

48. Our Approach Based on experimental modeling
No simplifying assumptions on the print-scan channel
Embeds significant volume of information
E.g., several hundred bits in 512x512 images
Rotation estimation using printer halftone information
No penalty for having invariance to rotation
Does not output halftone image directly
Bonus: Survives several other attacks
Gaussian or median filtering, heavy JPEG compression, scaling or aspect ratio change, or random bending

49. Experimental Setup Laser printers used Scanner
Lexmark, HP, Sharp
Scanner
Canon flatbed
Paper: Xerox recycled paper for printers
Variety of features were studied
Interesting trends observed in DFT magnitudes

50. Print-Scan Channel
Spectra of the original Image
Diff. in log DFT magnitudes of scanned and original image
Observation 1: Low frequency coefficients are preserved much better! Notice that most central part of right hand figure is closer to green.

51. Print-Scan Channel
Diff. in log DFT magnitudes of scanned and original image
Low Freq. Spectra of original Image
Observation 2: Error is high for low magnitude coefficients! Notice that all the dark blue points on the left correspond to the dark red points on the right.

52. Print-Scan Channel
Diff. in log DFT magnitudes of scanned and original image
Low Freq. Spectra of original Image
Observation 3: The coefficients that do not get washed out, see a gain of roughly unity!

53. Selective Embedding in Low Frequencies (SELF)
Take magnitude
2D DFT
DFT “Coefficients”
Log
Choose coefficients to hide
Threshold
Criterion
Image (NxN)
Hide using choice of scalar quantizer
exp., add phase and Inverse DFT

54. Coding Framework
The proposed technique is an Image-Adaptive technique.
Insertion-Deletion problem
Decoder uses the same criteria to guess embedding locations.
Might cause desynchronization.
Erasures at the Encoder
Use powerful erasures and error correcting codes in a novel fashion to counter the insertion-deletion problem.
Turbolike Repeat-Accumulate (RA) employed.

55. Undoing Rotation Laser printers use ordered halftoning.
In ordered halftoning, cells lie in a periodic array an angle of 45 degrees with the horizontal.
The halftone pattern can be captured by high resolution scanning.
Peaks in Fourier magnitude spectrum are used to determine halftone cell orientation.

56. Zoomed image and its DFT
Zoomed printed and scanned image
Fourier magnitude spectrum with primary peaks
Notice the peaks at 45 degrees

57. Rotated image Image rotated during scanning Fourier magnitude spectra
Notice that the peaks in the spectrum have also been rotated.

58. Example 1: Baboon image Original 512x512 Baboon image
Image with 475 bits hidden

59. Printed and scanned image – rotated during scanning process
Example 1 (cont.)
Printed and scanned image – rotated during scanning process
Automatically de-rotated image

60. All the hidden 475 bits decoded successfully
Example 1 (cont.)
Automatically de-rotated image
Image after automatic cropping
All the hidden 475 bits decoded successfully

61. Example 2: Man image Original 512x512 Man image
Image with 500 bits hidden

62. Printed and scanned image – rotated during scanning process
Example 2 (cont.)
Printed and scanned image – rotated during scanning process
Automatically de-rotated image

63. All the 500 hidden bits recovered successfully
Example 2 (cont.)
Automatically de-rotated image
Image after automatic cropping
All the 500 hidden bits recovered successfully

64. Results|More Image # bits hidden RA code rate # coeff. in band Peppers
250
1/4
870
Baboon
475
1/6
2450
Bridge
1/7
1560
Man
500
1/5
Couple
300

65. Automatic de-rotation Vs. Manual placing
Image
Number of bits hidden
Manual Placing
Automatic de-rotation
Peppers
225
250
Baboon
350
475
Bridge
200
Man
400
500
Couple
275
300

66. Robustness against other attacks
Images with the proposed SELF hiding scheme also survive following attacks :-
JPEG compression at QF = 10
Gaussian filtering
Median filtering
Rows and column removal, e.g.,
17 rows and 5 columns
5 rows and 17 columns
Stirmark random bending, to a limited extent

67. Conclusions Three Key Components:
SELF: Adaptive embedding strategy based on experimental modeling.
Use of powerful erasure and error correcting codes.
Rotation estimation using knowledge of printer halftoning algorithm.
Potential for many applications!

68. Print-Scan Resilience: Future Work
Mathematical modeling of the print-scan channel
Some results appearing at SPIE EI ‘05
Increasing hiding rate
Current hiding rates are small
Hiding capacity
Better understanding of the embedding capacity for this channel

69. Thank You 