Li et al.

EURASIP Journal on Image and Video Processing (2018) 2018:126

https://doi.org/10.1186/s13640-018-0358-7
EURASIP Journal on Image
and Video Processing

RESEARCH Open Access

Research on iris image encryption based on

deep learning
Xiulai Li1,2, Yirui Jiang1, Mingrui Chen1* and Fang Li3

Abstract
With the development of information technology, the demand for information security is increasing. For more
convenient and safer needs, the encryption technology based on biometrics has developed rapidly. Among them,
iris technology has become an important research object of information security research due to the stability of iris
characteristics and its difficulty in forgery. In this paper, the iris feature encryption technology based on the iris is
studied by using the method of deep learning as the feature classification method and the iris feature as the
research object. The simulation experiment is carried out by using the common iris database. The results show that
the method can greatly improve the consistency of iris encryption and improve the security of encryption and
decryption process.
Keywords: Image analysis, Iris, Deep learning, Image encryption

1 Introduction destruction. Image encryption algorithm aims at improving

With the development of technology, people have en- the security of image information, reducing the probability
tered the era of big data, Internet technology and com- of information leakage and destruction, and ensuring the
puter technology have developed rapidly, the popularity safe transmission of original information. In some special
of the network is very extensive, and information inter- circumstances, image data must be encrypted to transmit.
action technology is becoming more and more mature. For example, in medical image transmission, in order to
While people generally use the network to transmit in- protect patient privacy, these pictures need to be encrypted
formation, it also breeds many information security is- before they can be transmitted over the Internet. The secur-
sues. As the network enters various fields, the security of ity of image information involves a wide range of social sec-
information transmission increasingly affects the security tors, ranging from individuals to countries. Attacks such as
of individuals, enterprises, and even countries. malicious destruction and information theft by the crimi-
The image has a very good expression effect on the con- nals frequently occur, and image encryption technology is
tained information because of its visual characteristics. It is developing rapidly. How to improve image security,
widely used in information interaction, so many informa- anti-attack capability, and key transmission security is an
tion expressions are more favored for images. Owners with urgent problem to be solved [1].
important value images often use the Internet to achieve The rapid development of information technology, the
auctions or publish their own image information. This increasing need for information security transmission,
method overcomes the limitations of geographical condi- and the biometric-related technology are developing rap-
tions such as geography and is not only convenient and idly. Human biometrics refer to the information features
fast, but also reduces economic costs. However, in the produced by human tissue structure, including the iris
process of network transmission, the insecurity factors of [2–4], face [1, 5], fingerprint [6–8], voice [9–11], DNA
image information give the malicious attack the opportun- [12–14], and palm postures [15]. The biological charac-
ity to attack, and the original image information may be teristics of human body are widely used in identity recog-
attacked and lead to information leakage or information nition and other fields because of their uniqueness. Iris
recognition technology is used to extract the characteris-
* Correspondence: [email protected] tics of the iris. By comparing with the iris database image
1
Hainan University at Meilan, No. 58 People’ Rd. Meilan District, Haikou, China to calculate the similarity, the identity decision is finally
Full list of author information is available at the end of the article

© The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0
International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and
reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to
the Creative Commons license, and indicate if changes were made.
Li et al. EURASIP Journal on Image and Video Processing (2018) 2018:126 Page 2 of 10

realized. It belongs to the human biological feature extrac- Image encryption has higher security requirements.
tion technology and has a very excellent quality. Com- Iris features are more suitable for extracting feature from
pared with other human biological characteristics, the iris deep learning algorithm and applying it to image encryp-
has more excellent characteristics. It is more suitable for tion. Compared with other biometrics, it is more effect-
image encryption to improve the security and anti-attack ive to improve encryption security. The traditional iris
ability of the algorithm. Academic and business circles pay image encryption technology, the iris feature extraction
more attention to identity recognition technology based method which uses machine learning algorithm, presents
on iris feature extraction. It has a wide application pro- some problems of lack of feature learning and low effi-
spect and is gradually applied to some departments which ciency, such as image preprocessing trouble, high quality
need high security requirements, such as banking system, of iris image quality, and the need to repeat the iris
secrecy, and so on. image acquisition in the process of decryption in order
The target information contained in the image is filtered to achieve a correct decryption operation. In recent
by feature extraction to perform information processing years, deep learning has solved many unsolved problems
according to the user’s needs. The biggest disadvantage of in the field of artificial intelligence with its excellent fea-
traditional machine learning is that many parameters need ture learning ability and has achieved rapid develop-
to be set manually during the learning process. This short- ment. Research shows that deep learning is superior to
coming is especially noticeable when dealing with big data other learning algorithms in many fields, especially in
and high-dimensional complex data. Deep learning tech- feature learning on high-dimensional complex data and
nology belongs to machine learning [16–20], based on big data, and has been applied in commercial, economic,
artificial neural network for data feature extraction. The and government fields. The reason why deep learning
original information is expressed as a feature vector can be applied to image encryption technology is that it
through feature learning, input into a sub-learning system exhibits excellent learning ability for dealing with a large
(such as classifier, decision maker), and then the sample is amount of data, accurately extracts essential features,
classified or detected. Feature learning is a means of de- and satisfies the high security requirements of image
tecting and classifying input samples. Through feature encryption.
learning of input samples, feature extraction is automatic- In this paper, the deep learning algorithm is introduced
ally realized. to extract the iris image. Based on the original iris image
Deep learning uses a nonlinear neural network model to encryption algorithm, a new iris image encryption algo-
transform original data into higher-level abstract represen- rithm based on deep learning is proposed. The image en-
tations through nonlinear transformations. Deep learning cryption function is realized by training the sample. The
uses an artificial neural network structure. After multiple simulation experiments on the iris samples of the public
linear transformations and nonlinear transformations, iris database show that the proposed method can solve the
complex functions are constructed. The principle of sam- inconsistency of iris features and improve the confidential-
ple classification is to strengthen the ability to discrimin- ity of the encryption and decryption process.
ate data and weaken irrelevant factors by high-level
expression of original data, illustrated by an example: The 2 Proposed method
image is a matrix composed of many pixels. The first layer 2.1 Iris feature extraction
of the model expresses the features of the edge and pos- The structure of the human eye mainly includes the
ition of the image. The second layer will display the fea- pupil, the iris, and the sclera (as shown in the upper left
tures such as edges on the basis of the features extracted of Fig. 1). The pupil is an approximately circular black
by the first layer to detect the pattern, while the interfer- structure in the center of the eye. The white area around
ence item will be ignored. The third layer may further the eyeball is the sclera. The iris is located in the annular
splicing and combining the detected images to show region between the pupil and the sclera. The texture in-
the part that has obvious detection or classification formation is very rich. The details of the iris are mainly
help for the target. Thus, as the number of layers in- stripes, wrinkles, crypts, stains, and so on (as shown in
creases, the target to be detected is gradually formed. the lower left of Fig. 1). The best choice for biometric
Traditional machine learning is based on manual fea- technology is the iris. The reason is that the iris has the
ture setting and machine learning algorithm is used to advantages of highest uniqueness, high recognition ac-
screen. The core of deep learning is automatic feature curacy, high speed, strong biological activity, strong
setting in each layer. Model training is realized through anti-counterfeiting ability, and lifelong stability com-
continuous feedback, and features are constantly pared with other biological features.
learned from data. There are also some relevant re- It can be seen from the structure of the iris that the
search results on this aspect of research, which proves iris area of the human eye has a very rich texture feature.
that relevant research is feasible [21, 22]. The iris is formed in the stage of human embryo
Li et al. EURASIP Journal on Image and Video Processing (2018) 2018:126 Page 3 of 10

Fig. 1 Human eye structure and human eye sample

development, and the cell tissue grows randomly, result- extraction, and recognition. According to the above de-
ing in a series of differences in shape, color, and gray in scription of iris recognition technology, the iris recogni-
the iris area. These differences constitute iris texture fea- tion system framework is shown in Fig. 2.
tures. The different shapes of scars, different gray areas, As an important branch of pattern recognition, feature
and randomness of features make the difference between extraction and pattern classification of iris recognition are
iris very large. Texture is not affected by gene regulation, two important tasks. From the perspective of feature ex-
even though the iris texture of the right and left eyes of traction, the existing iris recognition algorithms can be
the same person varies greatly. After iris localization and roughly divided into three categories: based on phase
human eye image segmentation, how to effectively ex- method, zero crossing representation method, and texture
tract the texture features contained in the iris is the analysis method. From the perspective of the classification
most important step in the whole iris recognition sys- mechanism of iris samples, the existing algorithms can be
tem. Iris feature extraction refers to extracting the rich divided into three categories: the algorithm based on the
texture information contained in the iris as the feature distance classification, the method of using the correlation
code of the iris for subsequent matching recognition, analysis, and the classification algorithm through machine
and the distance measurement function determines the learning. There are few algorithms for classifying using
similarity between the two iris images. Iris recognition machine learning. The most popular is the use of learning
technology uses the invariance of the iris and the differ- vector quantization (LVQ) in the classifier.
ence of features to achieve individual identification. The iris contains rich features, and it is suitable for
Individual identity is determined by comparing the simi- recognition, encryption, and other fields because of
larities between features of the iris image. Image process- its excellent quality. How to extract high-efficiency
ing, pattern recognition, and other methods are used to and strong distinguishing features from iris images is
describe and match the characteristics of the iris to auto- a key part of iris application. Iris preprocessing
matically identify the identity. Typical iris recognition sys- mainly includes iris dryness, filtering, localization, and
tems include iris image acquisition, preprocessing, feature normalization.

Fig. 2 Iris recognition system framework

Li et al. EURASIP Journal on Image and Video Processing (2018) 2018:126 Page 4 of 10

There are many kinds of noise in the image, and the main The inner edge of the iris is positioned using the gray
purpose of image smoothing is to reduce noise. Their effect projection method. The center and radius of the pupil
on the amplitude and phase of the image signal is very are determined based on the gray scale distribution of
complex. Common noise includes additive noise, multi- the image. The gray scale distribution of the human eye
plicative noise, quantization noise, and “salt and salt noise.” image has certain characteristics. The gray scale distribu-
In the frequency domain, due to the high-frequency do- tion of the pupil, sclera, and iris is different, so the inner
main of the noise spectrum, various forms of low-pass fil- edge of the iris is defined by gray. In the imported
tering can be employed to reduce noise in the spatial image, find the sum of the pixels of the x and y axes, re-
domain. Image noise is often intertwined with signals, espe- spectively, and find the point with the smallest pixel. Fi-
cially multiplicative noise, but the basis of smoothing is de- nally, use the traverse method to find the boundary
tails such as edges and blurred contours, so smoothing point of the pupil.
noise removal is maintained as much as possible while The outer edge of the iris is positioned by the calculus
maintaining image detail. operator method. Based on the gray scale characteristics
Smoothing usually uses Gaussian template for digital of the human eye, the sclera is white, and the iris is dar-
image processing. It removes point mutations for fixed ker. In the image, the gray scale contrast is large, so the
point and several surrounding points to remove certain detection of the outer boundary of the iris needs to be
noise. The Gaussian template used in this paper is as treated as follows on every point on the image.
follows:
(1) According to the length and width of the input


image, the circle detection range of the center and

1 2 1


1

the corresponding radius is determined. According
2 4 2


16

to the circumferential parameters of the rough
1 2 1

location of the outer iris boundary, the range of the
value is narrowed and the positioning speed is
In the case where the noise is not obvious, the accelerated.
smoothness does not contribute much to the segmenta- (2) For all possible radii, count the gray scale average of
tion and matching of the image, but it becomes difficult the points on the circumference.
to find the edge point boundary due to the problem of (3) Find the gray scale gradient of two adjacent circles.
blurring. The purpose of sharpening the image is to (4) Find the radius of the gray gradient jump value.
make the edges and outline blurred images clear and the
details clearer. From a spectral point of view, the average The inner and outer edges of the iris are positioned as
or integral operation of the reverse operation (such as shown in Fig. 3. The left side of Fig. 3 is the inner edge,
differential operation) is usually used. The essence of the and the right side is the outer edge.
image blur is the attenuation of the high-frequency com-
ponent, which can make the image clear by the 2.2 Deep learning theory
high-pass filter. It should be noted that sharpening of Deep learning is a popular research direction in recent
the image must require a higher signal to noise ratio, years. Due to its unique network structure, its deepening
generally eliminating or reducing noise prior to sharpen- in academic and industrial fields is often referred to as
ing. The difference method is one of the common deep neural network. The deep learning network itself is
methods of image sharpening, and the difference an artificial neural network model, but it has improved
method is divided into two methods: first-order differen- in terms of network structure and learning algorithms
tial and second-order differential. Second-order differen- than traditional artificial neural networks.
tial operators are more sensitive to noise and tend to Deep learning is a new field of machine learning re-
enhance noise components, so smoothing operations search. By establishing an artificial neural network model,
usually use second-order differentiation. Although the it simulates the working mode of the human brain, so that
edge of the sharpened image is enhanced, the increase in it can make a correct judgment on the input information.
noise is more serious, and image smoothing can effect- After the deep learning network model is established, con-
ively reduce noise. tinuous learning is needed to optimize the training net-
The iris feature includes a lot of noise data, although the work model and improve the decision-making results. It is
noise reduction method can reduce the noise, but it can- similar to human beings who need continuous learning
not remove all the noise. For the eye sample selected in after birth to make more accurate judgments.
this paper (as shown in the right of Fig. 1), in order to bet- Deep learning is used to construct a network structure
ter extract the eye features, iris inner edge positioning and with multiple hidden layers. The entire network training
outer edge positioning method are used in this paper. has a large amount of training data, and the deep
Li et al. EURASIP Journal on Image and Video Processing (2018) 2018:126 Page 5 of 10

Fig. 3 Inner and outer edge location

network structure is used to extract features. Regardless (1) Calculate the error of the network layer output Op
of the human brain’s visual mechanism or its mathemat- and the ideal output Yp;
ical properties, the characteristics of the deep model are (2) The weight matrix is adjusted by inverse
far superior to the shallow model extraction feature, and propagation according to the algorithm such as
the use of this feature is very good for classification and gradient descent, and the parameters are
recognition. In this process, the depth of the multi-layer optimized.
hidden layer model is a method, and feature extraction
is the purpose of deep learning. CNN recognizes two-dimensional graphics with distor-
The convolutional neural network (CNN) is proposed tion invariance in the form of displacement and scaling.
by Yan Lecunn and his team. It belongs to a multi-layer The feature detection layer of CNN is learned from
neural network. The original CNN is mainly for the data training data, avoiding explicit feature extraction, and
processing of images and other data and has achieved learning from the training data implicitly when using
very advanced results. The CNN structure has three spe- CNN. In addition, the weights of neurons on the same
cial structures: convolution, downsampling, and weight feature mapped surfaces can be studied in parallel,
sharing. The network structure of CNN makes it very which is one of the advantages of CNN relative to the
few in terms of the number of parameters relative to neurons connected to each other. CNN has a unique ad-
other network structures. Each computing layer of the vantage in speech recognition and image processing. It
network consists of multiple feature maps, each of which has a special structure with local weight sharing, and the
is mapped into a plane with a higher distortion tolerance layout is closer to the actual biological neural network.
for the input samples in the recognition. The sharing of weight reduces the complexity of the net-
The CNN algorithm mainly includes two stages of for- work, especially for high-dimensional data. The input
ward propagation and back propagation, which are di- vector image can be directly input to the network fea-
vided into four steps: ture to avoid the complexity of data reconstruction in
The first stage, the forward transmission phase: feature extraction.

(1) Extract a sample (X, Yp) from the sample set and 2.3 Iris feature extraction method based on deep learning
enter X into the network (X, Y); Deep learning can learn by unsupervised, semi supervised,
(2) Calculate the corresponding actual output Op. or supervised methods of the original data, and extract the
advanced features contained in the information. It can be
In the forward propagation phase, information is con- used for pattern recognition, classification, and other
verted from the input layer to the output layer, and the scenes. Compared with the traditional feature extraction
network performs the training process after completing methods, it can produce better accuracy and achieve bet-
normal operations. In this process, the network is imple- ter application results. It is necessary to extract the fea-
menting the calculation (actually, the input weight matrix ture of the iris image by using the iris to generate the
is multiplied by each phase point, and the final output): image encryption key. However, the traditional feature
        extraction method is based on the image processing
Op ¼ F n … F 2 F 1 X p W ð1Þ W ð2Þ … W ðnÞ ð1Þ or shallow learning, and the operation is cumbersome
and the quality of feature extraction is poor. So the
The second stage, the backward propagation phase: feature extraction of iris image is carried out by deep
Li et al. EURASIP Journal on Image and Video Processing (2018) 2018:126 Page 6 of 10

learning. The feature extraction of iris image based (2) According to the error correction capability t, the
on deep learning refers to the calculation of iris generator polynomial g(x) on GF(2m) is obtained:
image by deep learning and extraction of the feature
matrix, which mainly includes iris image acquisition,
image preprocessing, feature depth learning, and fea-
ture extraction. Y
2t
g ð xÞ ¼ ðx−at Þ ð2Þ
The specific process of extracting iris features using i¼1
CNN can be described as the following steps:
The g(x) expansion formula is obtained according to
(1) Iris image acquisition of the human eye to obtain the domain operation rule simplification as follows:
an iris image;
(2) Iris image preprocessing to form an iris data set; X
2t
g ð xÞ ¼ ak ðiÞ xi ð3Þ
(3) Using the deep learning model CNN to learn the i¼0
collected iris image dataset, including parameter
initialization, model selection, and hyperparameter
selection, and using the inverse gradient descent
algorithm to train the model; (3) Encoding the information bit polynomial M(x)
(4) The output matrix of the deep learning model is according to the relation C(x) = M(x)g(x) to obtain
the extracted iris feature. the code word polynomial C(x).

The RS code belongs to the system linear block code,

2.4 Reed-Solomon error correcting code including the information bit and the supervised bit.
In the whole process of image acquisition, preprocess- The block code divides the continuous data bit stream
ing, and feature extraction of the iris, due to many into fixed-length groups, each group being further di-
factors such as the external environment, the iris fea- vided into m-bit symbols, usually taking 3 or 8 bits of
ture vector obtained by the two feature extractions data to form a symbol. The K symbols together form a
before and after cannot be exactly the same. Then, in source word, which is linearly encoded after being
the image encryption algorithm, it is necessary to en- encoded as a code word called an m-bit symbol block
sure that the encryption key used by the encryption code.
side and the decryption key used by the decryption
side are completely identical, so that the entire 2.5 Iris image encryption and decryption method based
process of encryption and decryption can be success- on deep learning
fully implemented. This creates a serious contradic- The iris image encryption algorithm based on deep
tion: the contradiction between the inconsistency of learning firstly performs normalization and other pre-
the iris feature vector and the strict correspondence processing on the collected iris image dataset and then
of the keys. The error correction code model can well uses the deep learning neural network model to extract
solve this contradiction. the features of the iris image. The extracted feature vec-
The Reed-Solomon error correcting code (referred to tor is used for key generation, and finally, the XOR oper-
as RS code) consists of two parts: the original code and ation is performed on the key and the pixel value of the
the check code. The original code contains the original original image.
data, and the check code is a data code generated after The encryption process can be described as the follow-
performing a certain regular operation on the original ing steps:
code. Due to the interference of external factors, when
the original code is different, the original code can be (1) Normalizing the iris dataset and using the iris
corrected by the check code, so that the original source dataset to train the deep learning neural network
code can be finally obtained. The coding algorithm steps model;
of the RS code are: (2) The encryption side collects the iris image and
inputs the trained deep learning model to extract
(1) Calculate the m query source polynomial table the feature vector V. The dimension of the feature
according to the relation n = 2m − 1 to obtain the vector V1 can be adjusted according to the adopted
original polynomial p(x) corresponding to GF(2m), image encryption algorithm;
thereby generating an extension field of GF(2m) and (3) Using the RS error correcting code to encode the
establishing a correspondence relationship between feature vector V, the encryption key Vk1 can be
the domain element d and the m vector. calculated;
Li et al. EURASIP Journal on Image and Video Processing (2018) 2018:126 Page 7 of 10

(4) Using the encryption key Vk1 and the pixel matrix 3 Experiments and results
gray value of the image matrix to perform XOR 3.1 Experimental dataset
operation to obtain an encrypted image, the entire In order to improve the credibility and predictability of
encryption process is completed. the algorithm, the experimental iris dataset uses the
CASIA iris database public version; CASIA is the first
The decryption process is performed after the image large-scale iris shared database built by the Institute of
is encrypted by the encrypting side, and after receiving Automation of the Chinese Academy of Sciences using
the image ciphertext and the RS error correcting code the self-developed iris image acquisition device. This
transmitted by the encrypting side, the reverse algo- paper uses the CASIA version 4.0. CASIA Iris Library
rithm of the encryption algorithm is used to realize the Version 4.0 has added three distinctive datasets
decryption of the ciphertext to obtain the plaintext. But ASIA-Iris-Distance, CASIA-Iris-Thousand, CASIA-Iris-
the algorithm in this paper is not a complete inverse Syn based on version 3.0 and was officially released in
transformation. The decryption process can be de- March 2009. By the end of November 2010, there were
scribed as the following steps: more than 3635 applicants from more than 100 coun-
tries and regions, including 522 domestic and 3113 for-
(1) Performing iris image acquisition on the eign countries, mainly in India (787), China (522), the
decryption side, inputting into the trained deep USA (290), Britain (114), Brazil (82), France (39), and
learning model, and realizing iris feature vector other countries. The database applicants include re-
extraction V2; searchers from internationally renowned universities
(2) Since V2 and V1 may have different values in some such as Harvard, MIT, CMU, and Cambridge, and other
dimensions, the RS error correction code is used to well-known international institutions, as well as R & D
correct the feature vector V2 to obtain the personnel from famous international enterprises such as
decryption key Vk2; Microsoft, Philips, BT (British Telecom), Sagem, and
(3) Using the decryption key Vk2 to encrypt the image other government institutions such as the American
matrix corresponding to the gray value of the pixel Standardization Technology Institute and the US Naval
to perform XOR operation, and get the decrypted Research Institute.
image, and then complete the whole process of The deep learning model for feature learning of the
decryption. iris uses CNN, and the dataset contains 400 images of
irises in total, 10 types, 40 images per class. The CNN
structure uses a five-layer network layer, the convolution
2.6 Evaluation method process is a 5 × 5 convolution filter, and the sampling
Iris image encryption is different from general image en- layer is a 2 × 2 pooling filter.
cryption algorithms and has different indicators in algo-
rithm evaluation. The evaluation indexes of iris image 3.2 Experimental simulation environment
encryption algorithm mainly include the following. In this paper, the data simulation is carried out in the
False acceptance rate and error rejection rate: The MATLAB 2014B environment.The data of each subject
false acceptance rate (FAR) is the probability that the is divided into test set and training set according to the
wrong key correctly decrypts the original image. The tenfold cross validation method. That is, the sample of
false rejection rate (FRR) is the probability that the the subjects was equally divided into ten parts, nine of
original image cannot be correctly decrypted with the which were training sets and one was a test set, which
correct key. were calculated ten times for each sample, and the final
System security: System security is the most import- results were averaged.
ant indicator to measure the encryption algorithm,
which restricts the feasibility of the encryption algo- 4 Discussion
rithm. It mainly includes indicators such as the prob- The feature extraction is performed on the iris as de-
ability of being attacked. The specific indicators and scribed above. First, the inner and outer edges of the iris
their implementation methods will be elaborated in the are positioned, and then the iris image is encrypted. The
experiments in the following chapters. There is a prob- encryption process is described in Fig. 4, and the result
ability that the attacked probability pointer will operate of Fig. 4 shows that the original inner and outer edge
on the encryption system and successfully decode the features of the iris are encrypted into a set of encrypted
plaintext. Unidirectionality refers to the ability to re- images that are unrecognizable in appearance.
cover plaintext from all aspects of the encryption Deep learning is a feature extraction method based on
process. Discrimination refers to the successful decryp- neural network. Training and testing precision after net-
tion performance of different irises. work training are an important indicator to measure the
Li et al. EURASIP Journal on Image and Video Processing (2018) 2018:126 Page 8 of 10

classification results. Figure 5 shows the training and test

accuracy comparison with the increase of sample
number.
It can be seen from the results in Fig. 5 that as the
number of samples increases, the accuracy also in-
creases. The reason is that when the number of samples
is too small, the extracted sample features are different,
and the learning mechanism does not function. As can
be seen from Fig. 5, when the number of samples is
above 600, the accuracy is significantly increased. For
test and training samples, the result of Fig. 5 shows that
the precision of the training sample is higher than that
of the test sample, because the training sample is the
feature set of the input learning machine fitting, so the
precision of the sample is higher than the test sample.
For the same person to perform feature extraction
twice, two sets of feature vectors are generated. If two Fig. 5 Prediction results for training and testing samples
set of feature vectors are defined as A1 and A2, then for
the random selected n group features, the difference Table 1 shows the difference results of the deep learn-
index between the description feature (the difference de- ing iris encryption features before and after encoding.
gree F value) calculation method is as follows: From the results of Table 1, the feature extraction differ-
ence of the same person collected in different cases is
0vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1 about 9%. The image encryption algorithm requires that
u 256
X
n uX   the encryption key must be consistent with the decryp-
@t A1ij −A2ij =ðA1=2 þ A2=2ÞA
tion key before it can be decipher successfully. It is said
i¼1 j¼1
f ¼ ð4Þ that the difference between the feature vectors is 0%, so
n that can successfully encrypt and decrypt as the key.
Therefore, Reed-Solomon error correction code is used
where i and j are the number of two sample set. to solve this problem. By comparison, the extracted

Fig. 4 Encryption of iris image

Li et al. EURASIP Journal on Image and Video Processing (2018) 2018:126 Page 9 of 10

Table 1 Comparison of iris characteristics Table 2 FAR and FRR for different thresholds
Number of groups 5 10 20 30 40 Threshold FAR (%) FRR (%)
Before coding 0.93 0.92 0.91 0.89 0.89 96 0.015 2.045
After encoding 0 0.01 0 0 0 95 0.014 1.882
94 0.007 1.402

256-dimensional feature vector is coded and corrected 93 0.013 2.206

by RS code, then the feature vector selects n groups ran- 92 0.003 1.043
domly for similarity matching and calculates the similar- 91 0.014 1.557
ity between the feature vectors according to the formula, 90 0 1.484
as shown in Table 1. The results were about 0. 89 0.005 1.972
Figure 6 shows the results of image encryption and de-
88 0 2.009
cryption using the iris.
In order to verify the feasibility of the algorithm, this 87 0 1.237
paper evaluates the algorithm from two aspects: decryp-
tion accuracy and security. The encryption algorithm re- probability of 2–428, so the deep learning algorithm is
quires that an illegal key cannot successfully decrypt the introduced to generate the key. It can greatly reduce the
ciphertext, so the FAR must be equal to zero. Due to chance that the key will be compromised, thus improv-
some unavoidable external interference during image ac- ing the security of encryption.
quisition, the iris image cannot be exactly the same.
Therefore, the generated encryption key and decryption 5 Conclusions
key cannot be completely consistent, and the legitimate In cryptography, the key is the key to successful encryp-
user cannot be decrypted absolutely successfully, and the tion and decryption. The security of information de-
repeated operation is required to achieve the successful pends on the security of the key. The traditional image
purpose. encryption algorithm cannot resist the attack of mali-
As can be seen from Table 2, at T = 92, FAR = 0.003%, cious key sharing and repudiation. If the key length is
FRR = 1.043%, and the RS error correction code is too large, it will be easy to lose and it will be difficult to
encoded by (440, 256). That is to say, the illegal key can- remember. Biometric encryption technology emerged as
not be successfully decrypted, and the legitimate key the times require, trying to solve the problem of poor se-
user has 1.043% need to perform two or more decryp- curity of key. The key is generated according to the bio-
tions to achieve successful decryption. metrics of the individual and then applied to the
After the encryption is completed, the ciphertext and corresponding image encryption algorithm to realize the
the RS error correction code are transmitted. According information encryption. The biological features that can
to this, the plaintext reply cannot be performed, indicat- be encrypted should meet the characteristics of uniqueness,
ing that the encryption algorithm has unidirectionality. stability, non-aggression, and so on. The iris not only satis-
When the key is attacked according to the RS error cor- fies the requirements mentioned above, but also has rich
rection code, only the guess can be made. When T = 92, feature information, strong anti-attack ability, and excellent
the key with the key length of 256 is attacked by the encryption potential. Iris image encryption has become an

Fig. 6 Encryption and decryption results

Li et al. EURASIP Journal on Image and Video Processing (2018) 2018:126 Page 10 of 10

important branch of image encryption and plays an import- Publisher’s Note

ant role in image encryption. Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
In this paper, the iris is taken as the research object, and
the image encryption and decryption process based on iris Author details
1
feature is realized. The iris feature extraction algorithm Hainan University at Meilan, No. 58 People’ Rd. Meilan District, Haikou,
China. 2Hainan Hairui ZhongChuang Technology Co. Ltd, Haikou, China. 3The
based on deep learning is established. The extracted fea- Maternal and Child Health Hospital of Hainan Province, Haikou, China.
tures are used for image encryption and decryption pro-
cessing, and the proposed algorithm is objectively Received: 15 August 2018 Accepted: 15 October 2018

evaluated. The simulation results of the public iris data-

base show that the proposed method can achieve image References
encryption. 1. M. Carcary, K. Renaud, S. Mclaughlin, et al., A framework for information
security governance and management. IT Prof. 18(2), 22–30 (2016)
2. X. Wu, N. Qi, K. Wang, et al., in Iihmsp '08 International Conference on
Abbreviations Intelligent Information Hiding and Multimedia Signal Processing. An iris
CNN: Convolutional neural network; DNA: Deoxyribonucleic acid; FAR: False cryptosystem for information security (IEEE, Elsevier Advanced Technology
acceptance rate; FRR: False rejection rate; LVQ: Learning vector quantization; Publications Oxford, 2008), pp. 1533–1536
RS code: Reed-Solomon code 3. J. Daugman, Information security technical report - prospectus: recognizing
people by their iris patterns. Inf. Secur. Tech Rep. 4, 29 (1999)
Acknowledgements 4. W. Abdul, A. Alzamil, H. Masri, et al., Fingerprint and iris template protection
The authors thank the editor and anonymous reviewers for their helpful for health information system access and security. J. Med. Imaging Health
comments and valuable suggestions. Inform. 7(6), 1302–1308 (2017)
5. S. Zhou, X. Lu, in Information Science and Management Engineering.
Fingerprint identification and its applications in information security fields
About the authors (IEEE, Shaanxi, 2010), pp. 97–99
Xiulai Li was born in Anqing, Anhui, People’s Republic of China, in 1992. He 6. M.M. Mahmoud Musleh, I.I. Ba, K.M.A. Nofal, et al., Improving information
received bachelor’s degree from Beijing Forestry University, People’s Republic security in e-banking by using biometric fingerprint: a case of major bank in
of China. Now, he studies in the College of Information Science and Malaysia. Int. J. Comput. Sci. Inf. Secur. 10(3), 1–4 (2012)
Technology, Hainan University. His research interest includes computational 7. J. Shang, H. Huang, in International Conference on Computer Science and Service
intelligence, cloud security, and information security. System. Shallow fingerprint identification information security technology in
Yirui Jiang was born in Zhumadian, Henan, People’s Republic of China, in the electronic commerce application (IEEE, 2011), pp. 1684–1687
1994. She received bachelor’s degree from Zhengzhou University, People’s 8. Ryu H. Information security attachment device for voice communication and
Republic of China. Now, she studies in the College of Information Science information security method for voice communication using the same. 2016
and Technology, Hainan University. Her research interest includes 9. L.A. Johnson, K.L. Dempsey, D. Bailey, et al., SP 800-128. Guide for security-
computational intelligence. focused configuration management of information systems. J. Dairy Sci.
Mingrui Chen was born in Chengmai, Hainan, People’s Republic of China, in 77(6), 1604–1617 (2011)
1960. She received bachelor’s degree from South China Normal University, 10. J.G. Chouhan, N.K. Singh, P.S. Modi, et al., Camera and voice control based location
Hainan, People’s Republic of China. Now, he works in the College of services and information security on android. J. Inf. Secur. 07(3), 195–205 (2016)
Information Science and Technology, Hainan University, His research 11. G. Cui, L. Qin, Y. Wang, et al., in IEEE International Workshop on Anti-
interests include software engineering, computer science, and information Counterfeiting, Security, Identification. Information security technology based
security. on DNA computing (IEEE, Xiamen, 2007), pp. 288–291
Fang Li was born in Hanchuan, Hubei, People’s Republic of China, in 1968. 12. G. Cui, L. Qin, Y. Wang, et al., Information Security Technology Based on DNA
She received bachelor’s degree from Tongji Medical University, Shanghai, Computing (2007), pp. 288–291
People’s Republic of China. Now, she works in The Maternal and Child Health 13. G. Cui, C. Li, H. Li, et al., in International Conference on Natural Computation.
Hospital of Hainan Province; her research interests include health care and DNA computing and its application to information security field (IEEE
health management during pregnancy and childbirth, Internet Med, and Big Computer Society, Tianjin, 2009), pp. 148–152
data on health. 14. Y. Mizuchi, Y. Hagiwara, A. Suzuki, et al., Monocular 3D Palm Posture Estimation
Based on Feature-Points Robust Against Finger Motion (2013), pp. 1014–1019
Funding 15. S. Bhilare, G. Jaswal, V. Kanhangad, et al. Single-sensor hand-vein multimodal
This work was supported by the funding project of A study on multiple biometric recognition using multiscale deep pyramidal approach[J]. Machine
mapping fusion Chinese and English automatic word alignment system Vision & Applications (2018) pp. 1-18.
based on multi-preprocessing mechanism takes the establishment of parallel 16. Y. Rao, J. Ni, in IEEE International Workshop on Information Forensics and
corpus on the Chinese-English translation website of Hainan tourism text as Security. A deep learning approach to detection of splicing and copy-move
an example (No. 20167238) and the funding project of Construction and ap- forgeries in images (IEEE, London, 2017), pp. 1–6
plication of female health system in Hainan (No. ZDYF2016136). 17. J. Ye, J. Ni, Y. Yi, Deep learning hierarchical representations for image
steganalysis. IEEE Trans. Inf. Forensics Secur. 12(11), 2545–2557 (2017)
18. M.E. Aminanto, K. Kim, in International Workshop on Information Security
Availability of data and materials Applications. Detecting impersonation attack in WiFi networks using deep
We can provide the data. learning approach (Springer, Cham, 2016), pp. 136–147
19. T.P. Le, Y. Aono, T. Hayashi, et al., in International Conference on Applications
Authors’ contributions and Techniques in Information Security. Privacy-preserving deep learning:
All authors take part in the discussion of the work described in this paper. revisited and enhanced (Springer, Singapore, 2017), pp. 100–110
The author XL wrote the first version of the paper. The authors XL and YJ 20. Z. Chen, D.O. Information, Face deep learning technology in the design and
did part experiments of the paper, and MC revised the paper in different implementation of the security in colleges and universities. J. Anyang Inst.
versions of the paper, respectively. All authors read and approved the final Technol. 16(6),70-75(2017)
manuscript. 21. M. Kumar, Y.H. Mao, Y.H. Wang, T.R. Qiu, C. Yang, W.P. Zhang, Fuzzy theoretic
approach to signals and systems: static systems. Inf. Sci. 418, 668–702 (2017)
22. W.P. Zhang, J.Z. Yang, Y.L. Fang, H.Y. Chen, Y.H. Mao, M. Kumar, Analytical fuzzy
Competing interests approach to biological data analysis. Saudi J. Biol. Sci. 24(3), 563–573 (2017)
The authors declare that they have no competing interests.