Professional Documents
Culture Documents
Abstract—Image hashing has become a major research area measures such as normalized Hamming distance, Euclidean
due to rapid growth of image alteration techniques that can distance, correlation coefficient, etc. [3]–[5]. The tamper de-
tamper digital images. The major concern of all image hashing tection capability means that perceptually different images
schemes is the selection of robust features. Local Binary Pattern
(LBP) is a technique that selects robust features for different must be separately distinguished and a minor manipulation
image applications. This paper presents a perceptual image in an image must change the corresponding hash values
hashing scheme by the utilization of Noise Resistant Local Binary significantly.
Pattern (NRLBP), a modified form of the LBP. The features One of the challenging tasks in a PIH scheme is the
of NRLBP are extracted from non-overlapping blocks of a selection of suitable features that are used in hash formation.
gray scale image. The NRLBP is combined with Singular Value
Decomposition (SVD) to provide good robustness characteristics The properties of the selected features have a significant
against a number of non-malicious distortions. Another major impact on the robustness and tamper detection capability of a
advantage of the proposed scheme is to detect localized tampered hashing scheme. Monga et al. [4] have used the non-negative
regions. Experimental results exhibit that the proposed scheme matrix factorization (NMF) vectors to calculate the hash of
has the capability to detect tampering as small as 3% of the image an image. The utilization of NMF provides good robustness
size and at the same time offers good robustness properties.
characteristics but may not be able to detect minute level of
Index Terms—Robust hash, perceptual image hashing, local tampering due to robustness of NMF vectors. Venkatesan et al.
binary pattern, singular value decomposition [6] have proposed an image hashing scheme using the wavelet
transform. This scheme is resilient to a number of content
I. I NTRODUCTION preserving manipulations, for example, JPEG compression,
With speedy development of image modification techniques, filtering, rotation, scaling and cropping. It is not clear whether
it has now become easy to tamper digital images. To counter small tampering within an image can be detected using this
this problem, Perceptual Image Hashing (PIH) techniques technique. Zhao et al. [7] uses global and local features to
can be used to validate image integrity. Conventionally, data generate the hash of an image. Zernike moments are used
integrity is provided by cryptographic hashing schemes. Cryp- as global features, while position and texture information of
tographic hashing schemes are however not appropriate for im- an image’s salient regions are used as local features. This
age integrity. The reason of non-compliance of cryptographic scheme can be used to detect localized forgery, including
hash functions with image integrity is its sensitivity to a single colour modification, however, the accuracy of tamper detection
bit change in the input data. This implies that two data streams greatly depends on the algorithm used for saliency detection.
with a single bit difference will produce a hash value with a The results reported in [7] reveals that this scheme requires
large Hamming distance. Images normally suffer from content further enhancement to detect minute level of tampering.
preserving manipulations and hence traditional cryptographic Recently, Davarzani et al. [5], [8] have proposed PIH
hash functions, such as SHA-1 [1], will not generate identical schemes using SVD based Center Symmetric Local Binary
hash values for similar visual images [2]. Pattern (CSLBP). The CSLBP is a modified form of the LBP
Feature extraction is the fundamental step in hash generation which was originally proposed by Ojala et al. [9] for texture
and selection of suitable features is a challenging task. The classification. The problem with CSLBP features is its weak
performance of any PIH scheme heavily depends upon the trade-off between robustness and tamper detection. For this
selected features; the more robust the features are, the more reason, it is not possible to detect small tampering in an image
efficient the hashing scheme would be. The performance of using CSLBP. Davarzani et al. [5], [8] notified in their work
a PIH scheme is evaluated under two distinct criterion, i.e., that the minimum tampered area must be 10% of the original
robustness and tamper detection capability. Robustness means image to successfully detect tampering. This, therefore limits
that perceptually similar images must have comparable hash the application of the schemes proposed in [5], [8] to detect
values. The comparison of hash values is performed by using tampering in an image area less than 10%. Chen et al. [10]
402
TABLE I
NRLBP H ISTOGRAM G ENERATION .
Algorithm 1
for All the pixels in an image
do
1. Compute 8-bit binary code (RN RLBP ) using Eq. 1.
2. Scan the RN RLBP code to calculate total number of
uncertain bits (q) in it.
3. Generate histogram
if q = 0 and the RN RLBP code is uniform;
Increment the corresponding uniform bin by 1.
else if q = 0 and the RN RLBP code is non-uniform; Fig. 2. SVD-NRLBP based Proposed Hashing Scheme.
Increment the 59th bin by 1.
else if q = 0
Compute the total number of the uniform codes (z) overlapping blocks of size 32× 32 pixels. The main advantage
that may be generated from the RN RLBP code. of using block-based approach is to localize tampered regions.
if z = 0 A Wiener filter is then applied to smooth each image block.
Increment the 59th bin by 1. The 32 × 32 blocks are further divided into 16 × 16 non-
else overlapping sub-blocks. The SVD transformation is applied to
Increment all uniform bins corresponding to the each sub-block to obtain the SVD decomposition given by Eq.
generated uniform codes by 1/z. 2.
(n) (n) (n)
end if SBm = Um Sm [Vm(n) ]T , (2)
end if
end for where, 1 ≤ m ≤ N , 1 ≤ n ≤ 4, N represents the total number
of image blocks of size 32 × 32 pixels, m is the image block
number under consideration and n represents a sub-block of
size 16 × 16 pixels within an image block m. Since each sub-
(n)
another case, when the RN RLBP code has some uncertain block SBm is of size 16 × 16 pixels, hence the size of U ,
bits in it, but it is not possible to generate even a single S and V is also 16 × 16. The matrices U and V contain
uniform code from it, then only the non-uniform bin (i.e. the orthonormal vectors, therefore, U U T = I, V V T = I. The
59th bin) is incremented by 1. For example, the RN RLBP matrix S is a diagonal matrix which contains singular values
code 11X0X100 cannot generate a uniform code, hence it is in the descending order of the corresponding sub-block. Let
(n) (n) (n) (n)
regarded as a non-uniform code and consequently the 59th umi and vmi be the ith column vectors of the Um and Vm
(n)
bin would be incremented by 1. The algorithm to generate matrices, respectively. The first column vector of the Um and
(n)
NRLBP code is given in Table I. To calculate an NRLBP code, Vm matrices of each sub-block are concatenated to form a
each pixel’s gray level is compared in a clockwise order with 16 × 8 SVD transformed block, Γm , given by Eq. 3.
it’s surrounding neighbours and the corresponding histogram (1) (2) (3) (4) (1) (2) (3) (4)
entries are increased. Γm = [um 1
, um 1
, um 1
, um 1
, vm 1
, vm 1
, vm 1
, vm 1
], (3)
Instead of directly applying NRLBP in the spatial domain, where superscripts of um and vm represents the sub-block
we apply NRLBP algorithm to the orthonormal vectors which (n)
number and can take the values 1, 2, 3 or 4. The matrices Um
are obtained by taking the SVD of the gray scale image under (n)
and Vm each has dimensions of 16×16 and at the same time
consideration. This increases the robustness of the hashing have orthonormal vectors. Only the first orthonormal vector of
scheme as demonstrated in [5]. The SVD decomposes a matrix dimension 16 × 1 corresponding to the highest singular value
into three different matrices termed as U , V and S. The diag- (n) (n)
is selected from each Um and Vm separately. Since the first
onal matrix S contains the singular values, while the matrices (n) (n)
column vector of each Um and Vm matrix corresponds to
U and V contain orthonormal vectors. In this paper, we show (n)
the highest singular value in Sm , hence the SVD transformed
that the fusion of SVD and NRLBP provide stable features that block would be less sensitive to non-malicious distortions [5].
are robust against content preserving manipulations and still As an example, the formation of the 16th SVD transformed
have the capability to distinguish minute malicious tampering block (m = 16) is illustrated in Fig. 3. The size of the SVD
in an image. transformed block (Γm ) is substantially reduced as compared
to the size of the original image block by utilizing the SVD
III. P ROPOSED A LGORITHM
decomposition. This dimensionality reduction approach further
The block diagram of the proposed scheme is shown in Fig. provides a robust structure of the block under consideration
2. An arbitrary input image is first converted into a gray scale and thus enhances the overall robustness of the proposed
image of size 256 × 256 pixels. It is then divided into non- hashing algorithm.
403
TABLE II
M INIMUM CORRELATION COEFFICIENT VALUE OF A BLOCK OUT OF ALL
THE 64 BLOCKS FOR THE C AMERAMAN IMAGE AND THE BABOON IMAGE
AFTER NON - MALICIOUS OPERATIONS .
404
(a) Cameraman Image. (b) Tampering is shown inside (c) Localized Tamper Detec-
the circle, Max. tampered area tion.
is approx. 3% of image.
(d) Baboon Image. (e) Tampering is shown inside (f) Localized Tamper Detection.
the circle, Max. tampered area
is approx. 3% of image.
Fig. 4. (a), (d) Original Image. (b), (e) Malicious tampered manipulations. (c), (f) Localized tamper detection.
the distorted version of the same image. The Cameraman im- value of r for all non-malicious content preserving distortions
age and the Baboon image were subjected to a number of con- is more than tr , indicating the robustness of the proposed
tent preserving distortions. These distortions are preservative scheme. Any value of r less than tr would indicate tampering.
Gaussian noise of mean (m) = 0 and variance (v) = 0.0005, D. ROC Curves
additive speckle noise of variance (Nv ) = 0.006, Gaussian
blurring of window size Fs = 11 × 11 and standard deviation To further evaluate the performance of the proposed scheme,
σ = 0.9, motion blurring with linear motion of camera by its ROC analysis is now presented. The term ROC stands for
L = 4 pixels and an angle of θ = 0o , gamma correction with Receiver Operating Characteristic and is a measure between
the value of γ = 1.5, image scaling by 30% to the original the false positive probability PF P and the false negative
image size and JPEG compression of 10%. The minimum probability PF N , while the threshold is varied. These two
value of r for a block out of all the 64 blocks for the probabilities are defined by Eqs 5 and 6, respectively [13].
Cameraman and Baboon images after applying non-malicious NAT
content preserving distortions are given in Table II. This result PF P = , (5)
NT
suggests that the value of tr should be less than 0.7213 to
NTA
positively authenticate these two images for the non-malicious PF N = . (6)
distortions under consideration. NA
In Eq. 5, NAT represents the total number of tampered image
C. Threshold Selection blocks detected as authentic, while N T represents the total
The minimum value of r for Cameraman and Baboon number of tampered image blocks. Similarly, in Eq. 6, NTA
images in case of non-malicious content preserving manipu- represents the total number of authentic image blocks detected
lations is 0.7213 while the value of r for malicious tampering as tampered and N A represents the total number of authentic
for both these images is 0.5237. It is quite evident that an image blocks.
impressive gap exists between malicious tampering and non- To calculate the false positive probability, the hash of each
malicious content preserving distortions. Hence, the selection of 32 × 32 block of the Cameraman image was compared with
of the threshold tr to separate the malicious tampering and the hash of the corresponding block located at the same spatial
non-malicious distortions is trivial. The value of tr can be position of the Baboon image. Since all the corresponding
set between 0.7213 and 0.5237. After doing experiments on blocks of both these images are visually dissimilar, ideally
a number of different images, it is concluded that the value there should be no false acceptance. The number of blocks
of tr = 0.65 is effective to discriminate malicious tampering that were positively authenticated by changing tr were noted
and non-malicious distortions. It is clear from Table II that the and the false positive probability was calculated using Eq.
405
(a) JPEG Compression. (b) Gaussian Blurring. (c) Gamma Correction.
5. To estimate the false negative probability, several distorted were observed to be quite robust against a number of content
versions of the Cameraman image were generated by applying preserving distortions as reported in this paper. The ROC
JPEG compression, Gaussian blurring, Gamma correction, analysis presented in Section IV reveals that a low false
Gaussian Noise, motion blurring and image scaling. The positive probability can be achieved for a low false negative
amount of distortion is same as shown in Table II. The hash of probability. As a future work, we aim to carry out the reported
each block of the original Cameraman image was compared experiments on a large data set with different malicious and
to the corresponding hash at the same spatial location of the non-malicious distortions to further gauge the effectiveness of
distorted image. The value of tr was changed and the number the proposed scheme. In addition, we would also work to make
of blocks not being authenticated were counted. The false the NRLBP-SVD feature key dependent. This would make the
negative probability was calculated using Eq. 6. The result of proposed hashing scheme both robust and secure.
ROC analysis is shown in Figure 5. The x-axis and the y-axis
represent the false positive probability and the false negative R EFERENCES
probability, respectively. It is promising to note that for low [1] S. William, “Cryptography and network security: principles and prac-
false positive probability, low false negative probability is tice,” Prentice-Hall, Inc, pp. 23–50, 1999.
achieved. For example, in case of JPEG compression, Gaussian [2] M. Wu, Y. Mao, and A. Swaminathan, “A signal processing and
randomization perspective of robust and secure image hashing,” in
blurring, Gamma correction and motion blurring, the values of Statistical Signal Processing, 2007. SSP’07. IEEE/SP 14th Workshop
PF P and PF N is approximately equal to 0.01. The values of on, pp. 166–170, IEEE, 2007.
PF P and PF N are a bit higher in case of Gaussian noise and [3] A. Swaminathan, Y. Mao, and M. Wu, “Robust and secure image
hashing,” Information Forensics and Security, IEEE Transactions on,
image scaling. vol. 1, no. 2, pp. 215–230, 2006.
[4] V. Monga and M. K. Mihçak, “Robust and secure image hashing via
V. C ONCLUSION AND F UTURE W ORK non-negative matrix factorizations.,” IEEE Transactions on Information
Forensics and Security, vol. 2, no. 3-1, pp. 376–390, 2007.
In this paper, an image hashing scheme is proposed that [5] R. Davarzani, S. Mozaffari, and K. Yaghmaie, “Perceptual image hash-
offers good robustness and at the same time is capable enough ing using center-symmetric local binary patterns,” Multimedia Tools and
to detect minute level of tampering. The fusion of NRLBP and Applications, pp. 1–29, 2015.
[6] R. Venkatesan, S.-M. Koon, M. H. Jakubowski, and P. Moulin, “Robust
SVD enabled to provide image features that are different for image hashing,” in Image Processing, 2000. Proceedings. 2000 Interna-
visually different images. On the other hand, these features tional Conference on, vol. 3, pp. 664–666, IEEE, 2000.
406
[7] Y. Zhao, S. Wang, X. Zhang, and H. Yao, “Robust hashing for image
authentication using zernike moments and local features,” Information
Forensics and Security, IEEE Transactions on, vol. 8, no. 1, pp. 55–63,
2013.
[8] R. Davarzani, S. Mozaffari, and K. Yaghmaie, “Image authentication
using lbp-based perceptual image hashing,” Journal of AI and Data
Mining, vol. 3, no. 1, pp. 21–30, 2015.
[9] T. Ojala, M. Pietikäinen, and D. Harwood, “A comparative study of
texture measures with classification based on featured distributions,”
Pattern recognition, vol. 29, no. 1, pp. 51–59, 1996.
[10] X. Chen, C. Qin, and P. Ji, “Perceptual image hashing using block
truncation coding and local binary pattern,” in 2015 Asia-Pacific Signal
and Information Processing Association Annual Summit and Conference
(APSIPA), pp. 856–859, IEEE, 2015.
[11] J. Ren, X. Jiang, and J. Yuan, “Noise-resistant local binary pattern with
an embedded error-correction mechanism,” Image Processing, IEEE
Transactions on, vol. 22, no. 10, pp. 4049–4060, 2013.
[12] M. Pietikäinen, A. Hadid, G. Zhao, and T. Ahonen, “Local binary
patterns for still images,” in Computer Vision Using Local Binary
Patterns, pp. 13–47, Springer, 2011.
[13] F. Ahmed, M. Y. Siyal, and V. U. Abbas, “A secure and robust hash-
based scheme for image authentication,” Signal Processing, vol. 90,
no. 5, pp. 1456–1470, 2010.
407