1. Applications of Data Hiding in Digital Images
Tutorial for The ISSPA’99,
Brisbane, Australia
August 22-25, 1999
Jessica Fridrich
Center for Intelligent Systems
SUNY Binghamton, Binghamton, NY , U.S.A,
and
Mission Research Corporation
1720 Randolph Rd. SE, Albuquerque, NM 87105, U.S.A
Fax/Ph: (607)

2. Outline Introduction to Data Hiding - History
- Motivation
- Definition
- Terminology
- Properties
Covert communication (steganography)
Digital watermarking (robust message embedding)
Watermarking for tamper detection and authentication
Attacks on hiding schemes
Open problems, challenges

3. Data Hiding in Digital Imagery
Relatively very young and fast growing
Well over 90% of all publications published in the last 6 years
Highly multidisciplinary field combining image and signal
processing with cryptography, communication theory,
coding theory, signal compression, and the theory of visual
perception
Tremendous interest from industry and military

4. Data Hiding - History
First techniques included invisible ink, secret writing using
chemicals, templates laid over text messages, microdots,
changing letter/word/line/paragraph spacing, changing fonts
Images, video, and audio files provide sufficient redundancy
for effective data hiding
Postscript files, PDF files, and HTML can also be used for
non-robust data hiding to a limited extent
Executable files, provide very little space for data hiding
Fonts

5. The Need for Data Hiding
Covert communication using images (secret message is
hidden in a carrier image)
Ownership of digital images, authentication, copyright
Data integrity, fraud detection, self-correcting images
Traitor-tracing (fingerprinting video-tapes)
Adding captions to images, additional information,
such as subtitles, to video, embedding subtitles or audio
tracks to video (video-in-video)
Intelligent browsers, automatic copyright information,
viewing a movie in a given rated version
Copy control (secondary protection for DVD)

6. Requirements Application Requirements Low High Covert communication
Copyright protection of images (authentication)
Fingerprinting (traitor-tracing)
Adding captions to images, additional information,
such as subtitles, to videos
Image integrity protection (fraud detection)
Copy control in DVD
Intelligent browsers, automatic copyright
information, viewing movies in given rated version
capacity
robustness
invisibility
Requirements
security
embedding complexity
detection complexity
Low
High

7. make data hiding possible
Redundancy
and
Irrelevancy
make data hiding possible
Information-theoretic
Removed by lossless
compression
Perceptual
Removed by lossy
compression
 2 gray
levels + =
 5 gray
levels + =
Original
 31 gray
levels + =

8. Data Hiding - Definition
Secret
message
Secret
message
Key
Carrier
document
Embedding
algorithm
Transmission
via network
Detector
Key
Relationship carrier - message
Who extracts the message? (source versus destination coding)
How many recipients are there?
Is the key a public knowledge or a shared secret?
Do we embed different messages into one carrier?
Embedding / detection bundled with a key in a tamper-proof hardware?
Is the speed of embedding / detection important?

9. Properties of hiding schemes
Robustness
The ability to extract hidden information after common image processing operations:
linear and nonlinear filters, lossy compression, contrast adjustment, recoloring,
resampling, scaling, rotation, noise adding, cropping, printing / copying / scanning, D/A
and A/D conversion, pixel permutation in small neighborhood, color quantization (as in
palette images), skipping rows / columns, adding rows / columns, frame swapping,
frame averaging (temporal averaging), etc.
Undetectability
Impossibility to prove the presence of a hidden message. This concept is inherently
tied to the statistical model of the carrier image. The ability to detect the presence does
not automatically imply the ability to read the hidden message. Undetectability should
not be mistaken for invisibility  a concept related to human perception.
Invisibility
Perceptual transparency. This concept is based on the properties of the human visual
system or the human audio system.
Security
The embedded information cannot be removed beyond reliable detection by targeted
attacks based on a full knowledge of the embedding algorithm and the detector
(except a secret key), and the knowledge of at least one carrier with hidden message.

10. The “Magic” Triangle Capacity Undetectability Robustness
There is a trade-off
between capacity,
invisibility, and robustness
Naïve steganography
Secure steganographic
techniques
Digital watermarking
Undetectability
Robustness
Additional factors:
Complexity of embedding / extraction
Security

11. Outline Introduction Covert communication (steganography)
Message hiding in RGB images
- Absolutely secure steganographic method
- LSB encoding
Message hiding in palette images
- Permuting the palette
- LSB encoding in the palette
- EZ Stego
- Improved EZ Stego
Digital watermarking (robust message embedding)
Watermarking for tamper detection and authentication
Attacks on watermarks
Open problems, challenges

12. Covert Communication
Purpose: To conceal the very presence of communication,
to make the communication invisible.
Encryption: To make the message unintelligible
Warden
Willie
Andy
Bob
Secret communication??!!
I just posted a picture of my
cat on my web page!

13. Covert Communication Secret message Encryption Unit Carrier Image
- Encryption and steganography
provide double protection
- Randomized message is easier
to hide
Encryption
Unit
Carrier
Image
Embedding
Algorithm
Modified
Carrier

14. Steganography for RGB images
Absolutely secure steganographic technique
Method:
Embed a small message (8 bits), by repeated scanning of
a cover image till a certain password-dependent message-
digest function returns the required 8-tuple of bits.
Comments:
Absolute secrecy tantamount to one time pad used in
cryptography
Guarantees correct noise distribution and undetectability.
Time consuming, very limited capacity, not applicable to
image carriers for which we only have one copy.

15. Steganography for RGB images
LSB Encoding (Least Significant Bit)
Method:
Replace the LSB of each pixel with the secret message
Pixels may be chosen randomly according to a secret key
Pixels may be chosen adaptively according to neighborhood
Message should always be encrypted
Comments:
The simplest and most common steganographic technique
Premise = changes to the least significant bit will be masked by
noise commonly present in digital images.
Color images provide more room for hiding messages
If more than one LSB is used, statistically detectable changes may
result
A provably secure method should introduce changes consistent
with the noise model

16. Steganography for palette images
LSB encoding cannot be directly applied to palette-based
images because new colors, that are not present in the palette,
would be created.
Two sources of palette images:
1. Color truncation + dithering of photographs
2. Computer generated images (fractals, cartoons, animations)
A secure steganographic method will produce modified carriers
compatible with the source

17. Possibilities Artifacts Hiding in the palette Hiding in the image data
Non-adaptive techniques
Adaptive techniques
Artifacts
Palette artifacts
Image data artifacts

18. Possible approaches Message hiding in the palette
Permuting palette entries
- Image is not modified
- Very limited capacity of log2(256!)=215 bytes
- Too fragile (resaving)
- Suspicious palette order is an artifact
LSB encoding in the palette
- Very limited capacity (at most 3256 bits)
- Palette artifacts?
Common disadvantage: Capacity is severely limited and
independent of the image size

19. Possible approaches
Message hiding in the image data - greedy techniques
Decrease color depth and expand
1. Collapse 256 colors  128 colors
2. Expand 128 colors  256 colors by including a close color
(e.g., flip the LSB of the blue channel)
3. Embed a binary message into the LSB of the blue channel
of randomly selected pixels
4. Read the message from the LSB of the blue channel
Alternatively
1. Decrease color depth to 32 colors and include all colors obtained
from LSB shuffling of all 32 colors (one color produces 23 new
colors)
2. Encode messages into the LSB of pixel colors
1 bpp
3 bpp

20. Possible approaches Message hiding in the image data Parity embedding
1. Assign parity to palette colors
2. Embed message bits as the parity of colors

21. Message:
Randomly chosen pixel with color C1
Find the color in the sorted palette C1
index = 30 =
Replace the LSB of the index to
color C1 with the message bit C2
The new index now points to a
neighboring color C2
Replace the index of the pixel in
the original image to point to the
new color C2.
Sorted palette
Critical assumption: Colors close in the luminance-sorted palette
are also close in the color space.

22. New approach using color parities
Message hiding in the image data
(1) For each message bit randomly select a pixel
(2) Calculate the set of the closest palette colors (in Euclidean norm)
The distance d between colors (R1G1B1) and (R2G2B2) is
d 2 = (R1–R2)2+ (G1–G2)2+ (B1–B2)2
(3) Find the closest color whose parity agrees with the message
bit. Parity of a color is defined as R+G+B mod 2.
(4) Change the index for the pixel to point to the new color.
To extract the secret message, pixels are selected using a key and
the secret message is simply read by extracting the parity bits of
the colors of selected pixels.
1 bpp

23. Message:
Randomly chosen pixel with color C1
Find the closest colors in the palette … …
Replace C1 with the closest color that has the same
parity as the message bit
Color parity of (R,G,B) = R+G+B mod 2.
Advantages over EZ Stego:
The total change to the image due to message embedding is always
smaller
We avoid occasionally large changes in color that are possible with
EZ Stego

24. Optimal parity assignment
Oblivious reading requirement:
The optimal parity assignment has to be reconstructable from
the modified image at the receiving end.
Optimal parity
Optimal parity
embed
Modified
carrier = =
Extract
message
message
Efficient algorithm for optimal parity assignment
Optimal parity depends only on the palette and does not
depend on the image content!
The optimal palette is also optimal for multiple-pixel embedding
The average decrease in the RMS error due to optimal palette parity
is about 25-35%.

25. Adaptive Steganography
Non-adaptive steganography = modifications due to message
embedding are uncorrelated with image features. Examples are LSB
encoding in randomly selected pixels, modulation of randomly
selected frequency bins in a fixed band, etc.
Adaptive steganography = modifications are correlated with the
image content (features).
- Pixels carrying message bits are selected adaptively
depending on the image
- Avoiding areas of uniform color
- Selecting pixels with large local standard deviation
Potential problem with message recovery: We have to be able to
extract the same set of message carrying pixels at the receiving
end from the modified image.

26. Computer
generated
Julia set
Large areas of uniform color
Internal structure of the image - it is a fractal Julia set
Fonts

27. Artifacts caused by non-adaptive methods
Artifacts around the Julia set.
Artifacts in the fonts.

28. Method 1: Adaptive block embedding
Message embedding
Divide the image into disjoint 33 blocks
Randomly choose blocks and evaluate some local
statistical quantity, such as standard deviation or
number of colors and decide whether or not a
message bit can be embedded (good vs. bad block)
If block is bad, skip it and do not insert message bit
If block is good, insert the bit into the block parity
If after embedding the block becomes bad, keep
the change but repeat the same message bit in the
next block
Message extraction
Generate the same random walk through the
image blocks
Read the parity from all good blocks

29. Limitations
Ultimately, image understanding is important for secure adaptive
steganography. A human can easily recognize that a pixel is actually
a dot above the letter "i" and must not be changed. However, it would
be very hard to write a computer program capable of making such
intelligent decisions in all possible cases.
Example of a difficult area for secure
adaptive message embedding - fonts on
a complex background

30. Embedding while dithering
True-color images are converted to palette images via
- color quantization
- dithering
Idea: To embed message bits while doing the dithering
Quantize
Dither
and
Embed
256 color
image
True color
image
Compute
palette
Increase color depth
by interpolating
Or start directly with
the true-color image

31. Embedding while dithering
1. Select a random collection of pixels that will carry message bits.
2. For non-message pixels use classical dither to the closest palette color
3. For message pixels dither to the closest color with the right parity.
Rounding error is added to the next pixel +
p11
p12+E11
Original 24-bit image
E11 = q11 - p11 Q Q
q11
q12
Dithered quantized image
Palette P = {q1, …, q256}
Non-message pixels: q is the closest palette color
Message pixels: q is the closest palette color
with the right parity Q:

32. Performance example Test image in JPEG format Original Non-adaptive
Embedding
while dithering

33. Outline Introduction, history, motivation, definition,
terminology, properties
Covert communication (steganography)
Digital watermarking (robust message embedding)
- Copyright protection of digital images (authentication)
- Fingerprinting (traitor-tracing)
- Adding captions to images, additional information to videos
- Methods for Robust Data Hiding (Watermarking)
- Image integrity protection (fraud detection)
- Copy control in DVD
Watermarking for tamper detection and authentication
Attacks on watermarks
Open problems, challenges