LOSSLESS DATA HIDING:

FUNDAMENTALS, ALGORITHMS AND APPLICATIONS

Yun Q. Shi1, Zhicheng Ni1, Dekun Zou1, Changyin Liang2 and Guorong Xuan3
1
New Jersey Institute of Technology, Newark, NJ, USA, shi@njit.edu
2
Shenzhen Polytechnic, Shenzhen, China
3
Tongji University, Shanghai, China

ABSTRACT retrieved for some legal or other considerations. The marking

techniques satisfying this requirement are referred to as
Recently, among various data hiding techniques, a new subset reversible, lossless, distortion-free, or invertible data hiding
called lossless data hiding has drawn tremendous interests. By techniques.
lossless data hiding, it is meant that the marked media can be Some lossless marking techniques have been reported in the
reversed to the original cover media without any distortion after literature over the last a few years. A classification of all existing
the hidden data are retrieved. After a careful study of all lossless algorithms into three categories is presented in Section 2. Some
data hiding algorithms published up to today, we classify the future research issues are discussed in Section 3. Conclusions are
existing algorithms into three categories: 1) Those developed for made in Section 4.
fragile authentication; 2) Those developed aiming at large
embedding capacity; 3) Those developed for semi-fragile 2. THREE CATEGORTIES OF EXISTING LOSSLESS
authentication. The mechanisms, merits, drawbacks and DATA HIDING ALGORTIHMS
applications of these algorithms are analyzed, and some future
research issues are addressed in this paper. In this section, all existing lossless data hiding algorithms
published in the literature up to today are classified into the
Key words—lossless data hiding, reversible data hiding, robust following three categories.
lossless data hiding, watermarking. .
2.1. Those for fragile authentication

1. FUNDAMENTALS The first several lossless data hiding algorithms belong to this
category. Since fragile authentication does not need much data to
Data hiding has recently been proposed as a promising technique be embedded in a cover media, the embedding capacity in this
for the purpose of information assurance, authentication, category is not large. Normally from 1k to 2k bits.
fingerprint, security, data mining, and copyright protection, etc. Barton’s patent in 2000 [4] may be the earliest one. His
By data hiding, pieces of information represented by some data algorithm was developed for authentication of digital media,
are hidden in a cover media. Many image data hiding algorithms including JPEG and MPEG coded image and videos. The main
have been proposed in the past several years. In most cases, the idea is to losslessly compress the bits to be overlayed and leave
cover media will experience some permanent distortion due to space for authentication bit-string. No specific performance
data hiding and cannot be inverted back to the original media. result has been reported.
From the literature, it is observed that most of the current Honsinger et al.’s patent in 2001 [5] is the second lossless
data hiding algorithms are not lossless. For instance, with the data hiding technique used for fragile authentication. Their
most popularly utilized spread-spectrum watermarking method is carried out in the image spatial domain by using
techniques, either in DCT domain [1] or block 8x8 DCT domain modulo 256 addition. In the embedding, Iw = (I + W ) mod 256,
[2], round-off error and/or truncation error may take place where Iw denotes marked image, I original image, W is the
during data embedding. As a result, there is no way to invert the payload comes from the hash function of the original image. In
marked media back to the original without distortion. With the the authentication side, the payload W can be reconstructed from
popularly used least significant bit-plane (LSB) embedding the marked image, then subtract the payload from the marked
method, the bits in the LSB are replaced according to the data to image to losslessly recover the original image. By using modulo
be embedded and the bit-replacement is not memorized. 256, over/underflow is avoided. However, the marked image
Consequently, the LSB method is not invertible. With another may suffer from the salt-and-pepper noise during possible
group of frequently used watermarking techniques, called grayscale flipping over between 0 and 255 in either direction due
quantization index modulation (QIM) [3], quantization error to modulo 256 addition. This issue will be addressed later in this
makes lossless data hiding impossible. paper.
In some applications, such as in the fields of law Fridrich’s group explored a deep research on lossless data
enforcement, medical and military image systems, in addition to hiding techniques and developed some algorithms. Their first
perceptual transparency it is desired to reverse the marked media algorithm [6] is in the spatial domain, which losslessly
back to the original cover media after the hidden data are compresses the bit-planes to leave room for data embedding. In

;‹,((( ,, ,6&$6

order to leave sufficient room for data embedding, it needs to pixel value to embed the data. In the image histogram, they first
compress a relatively high level bit-plane, which usually leads to find a zero point (no pixel assumes that gray value) and a peak
visual quality degradation. They also describe two reversible point (a maximum number of pixels assume that gray value).
data hiding techniques [7] for lossy compressed JPEG image. Then they move the histogram between zero point and peak
The first technique is based on lossless compression of biased point toward zero point by one unit and leave the histogram near
bit-streams derived from the quantized JPEG coefficients. The the peak point empty. Finally the histogram in peak point is
second technique modifies the quantization matrix to enable moved to its neighbor or kept intact to finish the embedding of
lossless embedding of one bit per DCT coefficient. In addition, ‘1’ or ‘0’. This algorithm has a quite large capacity (0.019 to
Fridrich’s group extended the idea of lossless authentication to 0.31 bpp) while keeping a very high visual quality for all images
MPEG-2 video [8]. (the PSNR of marked images versus original images is
guaranteed to be higher than 48 dB). This PSNR performance is
2.2 Those for high embedding capacity much higher than any other algorithms at the same data
embedding rate. .
All the above mentioned techniques aim at fragile Celik et al. [12] presented a high capacity, low distortion
authentication, instead of data hiding. As a result, the amount of reversible data hiding technique. In the embedding phase, the
hidden data is rather limited. Hence, Goljan et al. [9] presented a host signal is quantized and the residual is obtained. Then they
first lossless marking technique that is suitable for data adopt the CALIC lossless image compression algorithm, with
embedding. The details are as follows. The pixels in an image the quantized values as side information, to efficiently compress
are grouped into non-overlapped blocks, each consisting of a the quantization residuals to create high capacity for the payload
number of adjacent pixels. For instance, it could be a horizontal data. The compressed residual and the payload data are
block having four consecutive pixels. A discrimination function concatenated and embedded into the host signal via generalized-
is established to classify the blocks into three different LSB modification method. The experimental results show that
categories, Regular, Singular and Unusable. (The authors use the the PSNR and capacity are satisfying.
discrimination function to capture the smoothness of the groups.) Tian recently presented a new high capacity reversible data
An invertible operation F can be applied to groups. That is, it embedding algorithm in [13]. This algorithm employs two
can map between a pair of gray level values, resulting in F(R)=S, techniques, difference expansion and generalized least
F(S)=R, and F(U)=U. It is reversible since applying it to a gray significant bit embedding, to achieve a high embedding capacity,
level value twice produces the original gray level value. This while keep the distortion low. The reported results of ‘Lena’
invertible operation is hence called flipping F. The main idea for image is shown in Table 1. It seems until now this algorithm
lossless embedding is that they scan the image group-by-group reaches the highest capacity.
and losslessly compress the status of the image  the bit-stream
of R and S groups (the RS-vector) with the U groups simply Table 1. Payload size vs PSNR of embedded Lena image.
skipped  as overhead to leave room for data embedding. By Payload size
39k 82k 118k 172k 254k 369k
assigning a 1 to R and a 0 to S they embed one message bit in (bits)
each R or S group. If the message bit and the group type do not Bit rate(bpp) 0.15 0.32 0.46 0.67 0.99 1.44
match, the flipping operation F is applied to the group to obtain PSNR(dB) 44.2 41.6 37.7 34.8 29.4 24.0
a match. The data to be embedded consist of the overhead and
the watermark signal. While it is novel and successful in The main idea is described as follows. For a pair of pixel
reversible data hiding, the amount of data that can be hidden by values x and y , they first compute the integer average l and
this technique is still not large enough. From what is reported in
[9], the estimated capacity ranges from 0.011 to 0.092 bits per difference h of x and y , where h x  y . Then they shift h
pixel (bpp). This may not be high enough for some applications. to the left one unit and append information bit b in LSB. That
Another problem with the method is that when the capacity equals h' 2 u h  b . This method is called Difference
increases, the visual quality will drop severely. Expansion. Finally they compute the new x and y , based on
Xuan et al. proposed a high capacity lossless data hiding
the new difference values h' and the original integer average
technique based on the integer wavelet transform (IWT) [10].
IWT is used in the algorithm to ensure the lossless forward value l . In this way, the marked image is obtained. They only
transform and inverse transform. After IWT, the bias between embed information bits into the pixel pairs that cannot lead to
‘1’ and ‘0’ in the middle and high bit-planes of IWT coefficients overflow problem to overcome this overflow issue.
becomes much larger than that in spatial domain. Hence, those In summary, several lossless data hiding algorithms having
coefficient bit-planes can be losslessly compressed to leave a large embedding capacity have been developed. Method in [13]
large space for data embedding. Histogram modification is used may achieve the largest capacity, while method in [11] may
in this algorithm to prevent the over/underflow problem. The keeps the highest PSNR. If the embedding capacity of the
experiments show that the capacity can reach 0.057 to 0.36 bpp, improved method in [10] developed in IWT is comparable to
quite larger than the previous algorithms. While the PSNR of that in [13] is currently under investigation.
marked images are not very high due to histogram modification,
there is no any annoying artifact and the visual quality is 2.3. Those for semi-fragile authentication
satisfying. Further improvement has been made, resulting in both
higher embedding capacity and visual quality. For multimedia, content-based authentication makes more sense
Ni et al. [11] proposed a new lossless data hiding technique than representation-based authentication. The former is often
based on the histogram modification. This algorithm utilizes the called semi-fragile, while the latter fragile authentication. It is
zero or minimum points of the image histogram and modifies the because semi-fragile authentication allows some incidental

,,
modification, say, compression within a reasonable extent, to be “dense” that the original name of salt-pepper appears not to be
authentic. For this purpose, we need lossless data hiding appropriate.
algorithms robust to compression. That is, if the image has not Not only for medical images, the salt-pepper noise may be
been changed at all, it will be authentic. In addition, the original severe for color images as well. We have applied De
image will be able to recover. If the image has been compressed, Vleeschouwer et al’s method to eight JPEG2000 test color
the original image cannot be recovered, but the hidden data can images. Four images result in severe salt-pepper noise, while
still be recovered without error. other four less severe salt-pepper noise. Figure 3 presents an
De Vleeschouwer et al. [14] proposed a lossless data hiding example of severe case. This is a JPEG2000 test image. The data
algorithm based on patchwork theory, which has certain for authentication is embedded into the red color component. We
robustness against JPEG lossy compression. This is the only can observe severe salt-pepper noise, which manifests itself as
existing robust lossless data hiding algorithm against JPEG severe color distortion, i.e., about half of hair area becomes red
compression. That is, each bit of the message is associated with and most of palm area of right hand turns out to be green.
a group of pixels, e.g. a block in an image. Each group is equally
divided into two pseudo-random sets of pixels, i.e. zones A and
B. The histogram of each zone is mapped to a circle (positions
on the circle are indexed by the corresponding luminance). It is
observed that in most cases the vectors pointing to the mass
center of zones A and B are close to each other as shown in
Figure 1 (a) and (b).. Hence slight rotation of these vectors in
two opposite directions allows for embedding a bit. Embedding a
binary 1 is shown in Figure 1 (c) and (d). As to the pixel values,
rotations of the vectors correspond to luminance shifts. From
Figure 1, it is obvious, modulo 256 addition is used. Therefore
this algorithm suffers from the salt-and-pepper noise. More
investigation in this regard is presented in Section 3. Figure2. (a) Original Mpic1 image , (b) Marked image.

Figure 1: Data embedding diagram Figure 3. (a) Original Woman image (b) Marked image.
Secondly, the marked image does not have high enough
3. SOME FUTURE RESEARCH ISSUES PSNR. Table 2 contains test results for eight medical images.
The PSNR of marked image is as low as 26 dB (as 476
As pointed in Section 2, De Vleeschouwer et al’s method is the information bits are embedded in image of 512x512x8). Note
only existing lossless data hiding technique, which is robust to that the salt-pepper noise exists in each of eight marked medical
high quality JPEG compression. Hence it can be used for semi- images. When severe salt-and-pepper noise exists, the PSNR
fragile authentication. However, our extensive investigation may drop to below 10 dB.
reveals that it has some drawbacks that prevented it from
practical usages. . Table 2. Test results of eight medical images.
First of all, some marked images suffer from salt-and-
pepper noise because the algorithm utilizes modulo 256 addition. PSNR Robustness Salt-pepper
That is, in doing modulo 256 addition, a very bright pixel with a (dB) (bpp) noise
large gray value close to 256 will be possibly changed to a very Mpic 1 9.28 1.0 severe
dark pixel with a small gray value close to 0, and vise versa. The Mpic 2 4.73 2.0 severe
salt-pepper noise becomes severe for some images, where more Mpic 3 26.38 0.8 some
dark and bright pixels are contained. A typical example is
medical images. We tried eight medical images and found that Mpic 4 26.49 0.6 some
five images suffer from severe salt-pepper noise, and other three Mpic 5 26.49 0.6 some
result in some salt-pepper noise. Figure 2 gives such an example Mpic 6 5.60 1.6 severe
of severe noise. Note that the slat-pepper noise becomes so Mpic 7 9.64 0.8 severe
Mpic 8 5.93 2.8 severe

,,
Table 3 contains test results for eight JPEG2000 test images. The ACKNOWLEDGMENT
PSNR of marked image versus the original image is less than 30
dB (as 1412 information bits are embedded into a color image of The work on this paper was partially supported by the Digital
1536x1920x24). Note that the salt-pepper noise again exists in Data Embedding Technologies group of the Air Force Research
each of the eight test images. When severe salt-pepper noise Laboratory, Rome Research Site, Information Directorate, Rome
takes place, the PSNR can be as low as less than 20 dB. NY, under contract F30602-03-1-0264. The U.S. Government is
authorized to reproduce and distribute reprints for Governmental
purposes notwithstanding any copyright notation there on. The
Table 3. Test results with eight JP2000 color images.
views and conclusions contained herein are those of the authors
PSNR Robustness Salt-pepper and should not be interpreted as necessarily representing the
(dB) (bpp) noise official policies, either expressed or implied, of Air Force
 N1A 17.73 0.8 severe
Research Laboratory, or the U.S. Government.

N2A 17.73 2.2 severe REFERENCES

N3A 23.73 0.6 some
[1] I. J. Cox, J. Kilian, T. Leighton, and T. Shamoon, “Secure
N4A 19.67 1.2 some spread spectrum watermarking for multimedia,” in IEEE Trans.
N5A 17.28 1.2 severe on Image Processing, vol. 6. No. 12, pp. 1673-1687, Dec. 1997.
N6A 23.99 0.6 some [2] J. Huang and Y. Q. Shi, “An adaptive image watermarking
scheme based on visual masking,” Electronics Letters, 34 (8), pp.
N7A 20.66 1.4 some 748-750, 1998.
N8A 14.32 1.4 severe [3] B. Chen, G. W. Wornell, “Quantization index modulation: a
class of provably good methods for digital watermarking and
Thirdly, it is noted that in both Tables 2 and 3 there is an information embedding,” IEEE Transaction on Information
item called Robustness in the unit of bpp. This means that the Theory, vol. 47, no. 4, pp. 1423-1443, May 2001.
hidden data can be error-freely retrieved when an image [4] J. M. Barton, “Method and apparatus for embedding
compression ratio is above this quantity. We noticed, for authentication information within digital data,” U.S. Patent
instance, in Table 3, the JPEG2000 test image called N6A 5,646,997, 1997.
corresponds to 0.6 bpp, meaning that when N6A goes through a [5] C. W. Honsinger, P. Jones, M. Rabbani, and J. C. Stoffel,
compression, which is above 0.6 bpp then the hidden data by “Lossless recovery of an original image containing embedded
using De Vleeschouwer et al.’s method will be able to recover data,” US Patent: 6,278,791, 2001.
without error. However, this robustness is not strong enough for [6] J. Fridrich, M. Goljan and R. Du, “Invertible authentication,”
JPEG2000 compressed images. The reason is JPEG2000, due to Proc. SPIE, Security and Watermarking of Multimedia Contents,
its superiority, can compress an image to 0.0xx bpp while the pp. 197-208, San Jose, CA, January 2001.
compressed image is still in good quality for some applications. [7] J. Fridrich, M. Goljan and R. Du, “Invertible Authentication
Therefore, a robustness of 0.6 bpp is not enough for semi- Watermark for JPEG Images,” ITCC 2001, Las Vegas, Nevada,
authentication of JPEG2000 compressed images. pp. 223-27, April 2001.
It is noted that a method has been presented to eliminate [8] J. Fridrich, Rui Du, Lossless “Authentication of MPEG-2
salt-pepper noise in [14]. As the authors claimed [14], however, Video,” Proc. ICIP 2002, Rochester, NY
this method is not robust to JPEG compression. [9] M. Goljan, J. Fridrich, and R. Du, “Distortion-free data
Therefore, it is necessary to develop new robust lossless embedding,” Proceedings of 4th Information Hiding Workshop,
data hiding technologies that 1) do not use modulo 256 addition; pp. 27-41, Pittsburgh, PA, April 2001.
2) consequently, they will not generate the salt-pepper noises [10] G. Xuan, J. Zhu, J. Chen, Y. Q. Shi, Z. Ni, W. Su
and the marked images have high PSNR versus the original “Distortionless Data Hiding Based on Integer Wavelet
image; 3) the robust lossless data hiding techniques are robust Transform,” IEE journal, ELECTRONICS LETTERS, Volume
enough so that they can work for semi-fragile authentication of 38, No 25, pp.1646-1648, Dec.2002
JPEG2000 compressed images. [11] Z. Ni, Y. Q. Shi, N. Ansari and W. Su, “Reversible Data
Hiding,” IEEE International Symposium on Circuits and
Systems, Bangkok, Thailand, May 2003.
4. CONCLUSION [12] M. Celik, G. Sharma, A.M. Tekalp, E. Saber, “Reversible
data hiding,” in Proceedings of the International Conference on
This paper presents a thorough investigation on the development Image Processing 2002, Rochester, NY, September 2002.
of all existing lossless data hiding techniques. These techniques [13] J. Tian, “Reversible data embedding using a difference
are classified into three different categories. The mechanism, expansion,” IEEE Transaction on Circuits and Systems for
merits and drawbacks of each category are discussed. The future Video Technology, Vol. 13, No. 8, August 2003.
research issues on robust lossless data hiding algorithms, which [14] C. De Vleeschouwer, J. F. Delaigle and B. Macq, “Circular
may find wide applications in semi-fragile authentication of interpretation of bijective transformations in lossless
JPEG2000 compressed images, are addressed. It is expected that watermarking for media asset management,” IEEE Tran.
lossless data hiding may open a new door to link two groups of Multimedia, vol. 5, pp. 97-105, March 2003.
data: cover media data and embedded data. They will be applied
to information assurance such as authentication, secure medical
data system, intellectual property protection.

,,