Enhanced Biometric Template Protection Schemes for Securing Face Recognition in IoT Environment

With the increasing use of biometrics in Internet of Things (IoT)-based applications, it is essential to ensure that biometric-based authentication systems are secure. Biometric characteristics can be accessed by anyone, which poses a risk of unauthorized access to the system through spoofed biometric traits. Therefore, it is important to implement secure and efficient security schemes suitable for real-life applications, less computationally intensive, and invulnerable. This work presents a hybrid template protection scheme for secure face recognition in IoT-based environments, which integrates Cancelable Biometrics and Bio-Cryptography. Mainly, the proposed system involves two steps: 1) face recognition and 2) face biometric template protection. The face recognition includes face image preprocessing by the tree structure part model (TSPM), feature extraction by ensemble patch statistics (EPS) technique, and user classification by multiclass linear support vector machine (SVM). The template protection scheme includes cancelable biometric generation by modified FaceHashing and a Sliding-XOR (called S-XOR)-based novel Bio-Cryptographic technique. A user biometric-based key generation technique has been introduced for the employed Bio-Cryptography. Three benchmark facial databases, CVL, FEI, and FERET, have been used for the performance evaluation and security analysis. The proposed system achieves better accuracy for all the databases of 200-D cancelable feature vectors computed from the 500-D original feature vector. The modified FaceHashing and S-XOR method shows superiority over existing face recognition systems and template protection.


I. INTRODUCTION
T HE Internet of Things (IoT) [1] is the new revolution of modern access control, surveillance, smart city to smart homes, IoT-based health care systems, and even automated car driving, automated car parking, etc.In the case of access control and surveillance systems, networked IoT devices capture a person's biometrics and send them to the controlling system for decision making.The system decides whether the person will be allowed to enter or not.For the smart city, the IoT devices collect different data types through the respective sensors (e.g., a temperature sensor for weather forecasts) and provide information to the citizens.In the era of self-driving and autonomous vehicles, collision detection notification, automatic braking control, automatic parking, driver monitoring systems, and safety and security of cars need the use of a person's unique physiological or behavioral traits to achieve reliable, secure, and flexible access that is referred to as biometric authentication.Automated vehicle monitoring is among the most rising and demandable technologies [2].The safety and security of these autonomous vehicles are the first necessary and sufficient criteria for their performance.Vehicular automation includes mechatronics, which integrates mechanical, electrical, and electronics engineering systems, telecommunications, robotics, computer science, product engineering, and control systems.The semi-autonomous vehicle relies on automation for navigation, while the driver is in charge of the remaining responsibilities.The functionalities of the automated vehicle include collision detection notification, driver monitoring system, speed control adaptive headlamps, automatic braking control, automatic parking, blind-spot monitoring, automated night vision with pedestrian detection, and weather forecasting to activate climate mode.For all of the above cases, we need to store a massive amount of data as a future reference, and this data must be kept secure from attackers.Biometric-based recognition with flexible, robust, and high security increases the acceptability of IoT-based applications.
A biometric recognition system uses a person's biometric characteristics (face, iris, periocular, fingerprints, palm prints, voice, etc.) to recognize him/her automatically.Biometrics is the only reliable solution for Govt.or public sector undertakings (PSUs), e-healthcare [3], e-finance, border security control, immigration counter, banking applications, entry/access control, and many more.Factors, such as spoofing attacks, expensiveness, security, and privacy of the stored template, are the major obstacles to biometric-based systems.Among the biometric traits, the face is the most interactive because it can be captured in tangible and intangible modes.In a pandemic situation like COVID-19 (where there is a possibility of infection through touch), capturing biometrics in intangible mode is compulsory; the face is the best biometric trait for this.Face recognition is essential to the safety and security of autonomous vehicle monitoring systems.Facial recognition can improve driver safety by adjusting airbag intensity and position depending on the driver's head and body posture.But, the face is open to all on social media, or anyone can capture it without the user's consent and cause a spoof attack.Hence, it is most risky than any other biometric characteristic.When a registered user's biometric template is disclosed to hackers, it affects the security system by replay attack, spoof attack [4], reconstructed image attack [5] from the transformed template.Apart from these issues, biometric recognition system also suffers from some other attacks discussed in [6], such as 1) dictionary attack; 2) brute-force attack; 3) database attack; 4) channel attack; and 5) false acceptance attack.Biometric traits, being irreplaceable, cannot be discarded and re-enrolled using the same if disclosed.Moreover, there is a chance of cross-matching templates across different databases enrolled using the same biometric trait and detecting whether a user is registered in several unrelated applications.This creates serious privacy concerns for persons registered in the biometric system.In contrast with the above challenges, there are three solutions to protect biometric templates: 1) cancelable biometrics; 2) biometric cryptosystem; and 3) image transformation.The first two ways are featurelevel, and the third is an image-level template protection scheme.This work employed a combination of cancelable biometrics and a biometric cryptosystem.The main objective of this work is to implement a biometric recognition system in an encrypted cancelable domain for use in an IoT environment to preserve the original biometric feature in offline mode as future reference and keep the encrypted cancelable feature online for identification or verification purposes.
Objectives: The objectives of this article are as follows.
1) Implementation of a hybrid template protection scheme for a biometric recognition system in an IoT environment.2) Implementation of a novel Bio-Cryptography and a key generation technique.Contributions: The contributions of this article include: 1) a hybrid face template protection scheme using cancelable biometrics and Bio-Cryptography for a face biometric recognition system in an IoT environment has been employed and 2) a novel Sliding-XOR (S-XOR)-based Bio-Cryptography and a key generation technique from user biometrics has been introduced.
The organization of this article is as follows.Section II discusses the related works of this system.Section IV describes the implementation of the proposed methodology.Section V demonstrates experimental results and discussions.Section VI presents the conclusion and future scope of the system.

II. RELATED WORKS
This section investigates the papers describing face recognition systems (FRSs) with template protection schemes, security of IoT environments, and cryptographic algorithms.Qin et al. [7] introduced a method for face recognition by integrating Gabor wavelet and linear discriminant analysis.Sardar et al. [8] proposed a cancelable FRS (CFRS) using the FaceHashing technique that worked for both verification identification purposes.The deep-learning-based FRSs with partial facial images are experimented in [9], [10], and [11].Li et al. [12] incorporated deep learning for IoT into the edge computing environment to improve network performance and ensure user privacy when uploading data.Xiao et al. [13] proposed machine learning-based security techniques for IoT applications.They investigated the attack model for IoT systems and surveyed the solutions to those attacks on IoT security systems based on supervised, unsupervised, and reinforcement machine learning techniques.Alharbi et al. [14] proposed a security system based on FOG computing in IoT systems.The proposed FOCUS system uses a virtual private network (VPN) to secure the access channel to the IoT devices.Popescu [15] proposed a secure protocol for payment systems using ElGamal algorithms, which combines the ElGamal encryption scheme, ElGamal blind signature, and ElGamal signature scheme.Shahzadi et al. [16] employed an enhanced Rivest Cipher version 5 (RC5) encryption algorithm in a remote health monitoring system for the security and integrity of clinical images.Rachmawati et al. [17] employed the ElGamal algorithm for image compression and security.The FRS with a hybrid template protection scheme for cyberphysical-social services has been implemented in [18].
Imran et al. [19] proposed the ElGamal algorithm for the encryption-decryption of speech signals.Sardar et al. [20] had proposed a palmprint recognition using statistical patch-based feature representation technique and also introduced a noninvertible BioCryptosystems to preserve biometric templates.Dissanayake [21] improved the ElGamal to achieve better security which avoids plain text attacks.Yousif et al. [22] proposed an image encryption technique combining scanning, ElGamal algorithm, and chaotic systems.They used a zigzag and spiral scanning technique to construct a permuted image.These images are encrypted by the ElGamal encryption method, and finally, chaotic systems are used to scramble the pixel locations.Cahyono et al. [23] designed an FRS for employee presence using the Facenet algorithm, fivefold crossvalidation (CV) on support vector machine (SVM) classifier, and according to their investigation, the system achieves 100% accuracy for the FaceNet model while the Openface model achieves only 93.33% accuracy.Medapati et al. [24] proposed an IoT-based FRS for the smart cities safety management.Masud et al. [25] proposed an FRS for the cloud environments using a tree-based deep learning model.They achieved 95.84%, 99.19%, and 98.65% accuracy for LFW, ORL, and FEI databases, respectively.Rukhiran et al. [26] investigated on the performance of face recognition using IoT-based solutions to measure the impact of environmental conditions.

A. ElGamal Method
The ElGamal algorithm [27] is the continuation of the Deffie-Hellman key exchange method [28].It is a public-key cryptosystem based on the difficulty of computing discrete logarithms in a cyclic group.This algorithm consists of four steps, and these are shown in Fig. 1.The security of this algorithm depends on the computational difficulty of discrete logs in a big prime modulus.

B. RC5 Method
RC5 is a symmetric key block cipher encryption algorithm formulated by Ronald Rivest in 1994 [29].Being XOR and shift-based operation, it is faster and consumes less memory space.Each instance of RC5 is defined as w/r/b, where w, r, and b denote the word size (in bits) of input plain text, the number of rounds, and the key size (in bytes).Word size (w) can be 16, 32, or 64 bits, possible rounds (r) can be 0-255, and key sizes (b) range from 0 to 255 bytes.The block size of input plain text may be 32, 64, or 128 bits in size because RC5 addresses two-word blocks simultaneously.This algorithm has the following steps.
e) Subkey Integration: Mixing user's secret key with sub key S and a temporary array A as follows: ) mod m, j = (j + 1) mod n. 2) Encryption: The input plain text block is split into two wbit registers, X and Y, for the encryption operation.Then, two subkeys, S(1) and S(2), are generated and added with X and Y, respectively.The added results are stored in X and Y.Then, we have performed 1) X = X ⊕ Y; 2) cyclic left shift updated value of X by Y bits; and 3) add S(2 * i + 1) with the previous value, respectively, then we obtained the final value of X, and these operations are repeated r (number of rounds) times.Similarly, we computed the final value of Y. Finally, the encrypted text block is obtained by combining X and Y results.The encryption process is shown in Algorithm 1.

3) Decryption:
The encrypted text is split into two registers, X and Y, each with w bits length.Then repeatedly perform, 1) (Y −S[2×i+2]); 2) cyclic right shift Y by X bits; and 3) XOR operation between previous result and X, then result is stored into Y.Similarly, the computed value of X.Finally, subtract subkeys S(2) and S(1) from Y and X, respectively.The combined results of X and Y generate the decrypted text block.The decryption process is shown in Algorithm 2.

C. RSA Method
RSA [30] is a public-key encryption algorithm based on the multiplication of two long prime numbers p, q, i.e., N = p.q.The steps of RSA algorithm are shown in Fig. 2.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

IV. PROPOSED METHODOLOGY
In this work, we have evaluated the efficiency of cryptographic algorithms for face recognition within an encrypted domain for identifying or verifying individuals.The proposed system consists of 1) preprocessing; 2) feature extraction; 3) classification; 4) cancelable biometrics; and 5) Bio-Cryptography.These components are discussed below.

A. Image Preprocessing
In image preprocessing, all the challenging facial issues have been considered, including frontal and profile face, expression, off-angle, rotation, illumination, accessories, motion blur, etc.The face region has been extracted from the body silhouette.Then, the tree structure part model (TSPM) [31] has been used to compute 68 frontal face landmark points and 39 profile face landmarks points (−45 • to 45 • ) (Fig. 4).The 68 landmarks are the facial coordinate points in the eyebrows, eyelashes, eyes, pick of the nose, tip of the nose, nostril, lower and upper lip, jaw, and mouth contour.The landmarks are coordinate points (x abscissa, y ordinate).The x-abscissas of these coordinates represent the columns, while the y-coordinates represent the rows in digital image form.We apply the min-max method to the x-abscissa and yordinate of the coordinate points we computed (x-min, x-max) and (y-min, y-max), respectively.Based on these coordinate points, the four corner points (x-min, y-min), (x-min, y-max), (x-max, y-min), and (x-max, y-max) are combined to obtain the facial region.The 39 landmark points consist of one eyebrow, one eye, the nose pick, the nose's tip, one nostril, half of the lower and upper lip, half of the jaw, and half of the mouth contour.Then, we computed (x-min, x-max) and (y-min, ymax) for the x-abscissa and y-ordinate of the 39 coordinate points.Finally, combining the four corner points (x-min, ymin), (x-min, y-max), (x-max, y-min), and (x-max, y-max) from 39 coordinate points, we obtain the profile facial region.
After that, we estimated four corner points on each input image considering the pixel location of the TSPM-computed landmark points.Then, we extracted the face region using those corner points from the original image.

B. Feature Extraction
In this step, the discriminant features are computed from the preprocessed face regions F [Fig. 4(e)].In this work, four different feature representation approaches have been employed, such as 1) features from entire image F ; 2) splitting original image F into two equal segments horizontally, i.e., <P H1 , P H2 >; 3) splitting original image F into two equal segments vertically, i.e., <P V1 , P V2 >; and 4) splitting original image F into four square segments, i.e., <P S1 , P S2 , P S3 , P S4 >.Then, the ensemble patch statistics (EPS)-based feature extraction technique has been used [Fig.5].To extract more useful local information, a small patch ω 25×25 has been selected over segmented parts of F , which slides vertically followed by horizontally to define texture primitive or texels which give a better representation of F .The features extracted from these small patches form normalized vectors called texels t n 2 ×1 i (n = 25).Then, these local features are organized to get global representation for the face region.Then from each F , N number of texel vectors are generated, i.e., T i = [t 1 , t 2 , . . ., t N ].Now, if M number of image samples are selected for training, then we obtain N × M number of texels, i.e., {T 1 , T 2 , . . ., T M }.Then, we applied the K-means clustering to group these N × M texels which computes a corpus C ∈ R n 2 ×K , K(here, K = 250) (K << M) refers to the distinct texels in C .Finally, the computed corpus C and texels [t 1 , t 2 , . . ., t N ] undergo for feature computation.
The texels from each image or its segments, as well as the corpus C , are now considered while extracting features.Then, initialize feature f F (1 • • • K) ← 0, where K is the number of texels in C .Now, using α j = dist(t i , C i ), ("dist" refers to the Euclidean distance) determine the K-most similarity of each text on t n 2 ×1 ∈ T in C .The values of the K-most β j are then be updated in the appropriate position of f (j) such that f (j) = f (j) + β j where β j = exp{−(α j /n 2 )}.

C. Classification
A multiclass linear SVM with a fivefold CV technique has been employed to classify the subjects.During authentication, the identification and verification performances are investigated.Different ratios of training-testing samples are tested during the classification task.For identification purposes, the S different scores are obtained by comparing each sample with the S prototype of the S enrolled individuals.The scores are then arranged in decreasing order, and each score is given a rank.The highest score is given rank 1, the secondhighest score is given rank 2, and so on.Therefore, the accurate recognition is determined by how many times each subject's rank 1 matches correctly over the total number of subjects with its actual class membership to its class is called correct recognition rate (CRR).This CRR % refers to the identification performance.We computed an equal error rate (EER) for verification performance, assessed based on the true-positive and false-positive rates derived by the proposed system for the number of individuals enrolled.

D. Proposed Hybrid Template Protection Scheme
The proposed HTPS is the integration of two different approaches: 1) Cancelable Biometrics and 2) Bio-Cryptography.Cancelable biometrics is employed to achieve reusability, noninvertibility, unlinkability, and performance preservation properties of the employed biometrics.Finally, a robust and reliable Bio-Cryptography called S-XOR has been employed on the generated cancelable biometrics to store biometric features in an encrypted form in the database.
1) Proposed Cancelable Biometrics: In the proposed cancelable biometrics, we employed the FaceHashing technique to protect the feature vectors where the existing BioHashing technique has been enhanced to generate more secure cancelable biometrics.The proposed modified FaceHashing technique consists of three steps called "CFRS" level-1, level-2, and level-3, i.e., CFRS 1 (1), CFRS 2 (2), and CFRS 3 (3), respectively.At level-1 FaceHashing, a user token (t subject ) has been employed to generate a random matrix.Then, this random matrix is normalized to R ∈ R D ×m (where D is the dimension of the original feature vector, m is the dimension of the projected feature vector and D ≫ m) by "Gram-Schmidth Orthogonalization" scheme which is projected on each column of original feature vector f F ∈ R 1×D to compute x F ∈ R 1×m .This x F ∈ R 1×m is then quantized into X F ∈ {0, 1} 1×m called "FaceCode" and used for the verification purposes.This CFRS 1 shown in ( 1) is the existing BioHashing technique Some existing cancelable biometric systems use this CFRS1 approach to authenticate users using their assigned tokens.
This CFRS 1 technique is used as user authentication in several existing cancelable biometric systems with their assigned token.Moreover, this CFRS 1 is less secure.To implement it for identification mode biometric system and to enhance the security, we upgraded the CFRS 1 from (1) to CFRS 2 in (2), i.e., X F ∈ {0, 1} 1×m to Y F ∈ Z 1×m (i.e., each element of X F is decimal) which can be used for both verification and identification of a person.We considered the several feature dimensions, such as m = {100, 200, 300} for X F ∈ {0, 1} 1×m and Y F ∈ Z 1×m in both CFRS 1 and CFRS 2 , respectively Now, to enhance both performance and security levels, CFRS 2 is further extended to CFRS 3 applying permutation operation (π t 1 ) based on the token t 1 = t subject + t system , where t system is the system assigned token.Further, we applied the permutation operation (π t 2 ) based on the token t 2 = t subject + t system , where t system is another system assigned token.This CFRS 3 has been formulated in (3), and it is the proposed FaceHashing method.Finally, the cancelable features vector is transformed into a decimal vector for further operations 2) Proposed Bio-Cryptography: The proposed S-XOR is based on bit-wise XOR operation between elements of a biometric feature vector and a secret key.After each XOR operation, the secret key slides one position toward the left or right.Let a feature vector U = [u 1 , u 2 , u 3 , . . ., u n ] is to be encrypted by two secret keys K 1 and K 2 using the S-XOR approach.The key K 1 is transformed into a 256-bit binary, and K 2 is transformed into a 2048-bit binary.Now split K 1 into some parts (i.e., subkeys) of equal bit length, let us say 128-bit.Similarly, split K 2 into 128-bit length subkeys.Finally, perform bit-wise XOR operation by each subkeys of K 1 and K 2 sequentially with binary value of each element (u i ) 2 ∈ U, i = 1, 2, . . ., n.If the S-XOR moves from the most significant bit (MSB) to the least significant bit (LSB), then after each XOR operation by a subkey, the key moves one bit to the LSB.During decryption, S-XOR is performed in the reverse direction by subkeys of K 2 then K 1 sequentially.Since the S-XOR scheme uses the same keys for encryption and Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

Algorithm 3 Key Generation From User Biometric
Input: Preprocessed face image samples f F Output: Prime numbers P, Q, R, S primeList(i)=D(i) 14: end for 15: Choose any four prime numbers P, Q, R, S from the primeList.16: If the primeList does not have four prime numbers then apply bit shuffling and repeat step 4 to 12.
17: Compute λ = (P − 1) decryption, hence it is called a symmetric key cryptographic technique which consists of three steps-1) key generation; 2) encryption; and 3) decryption and these are discussed below.a) Key generation: For the proposed Bio-Cryptography, we used a novel key generation technique from the user biometrics.Preprocessed images are first enlarged to a dimension of 512 × 512, then binarized to produce 512-bit binary bit-streams.After that, these bit-streams of each row are transformed to decimal values.Finally, we applied the primality test to the computed decimal numbers and then four prime numbers are selected.The reason for selecting image sizes 512 × 512 is to generate the 512-bit prime number.Here, image height may be more or less than 512 pixels but image width must be 512 pixels because row-wise bitstreams are considered to generate prime numbers.To generate an encryption/decryption key, we considered four 512-bit random prime numbers (P, Q, R, and S) from the enrolled face biometrics.Then, a bit-wise XOR operation is performed among all prime numbers to compute a key λ in such a way that λ = and then the modulus N is computed by multiplying these four prime numbers, i.e., N = P × Q × R × S.This key generation process is summarized in Algorithm 3. Here, the numeric values (1-4) have been used to make the diagram flow more understandable.In step 1, preprocessed grayscale images are binarized.In step 2, if the LSB of the binarized matrix is 0, then replace that LSB 0 with 1.In step 3, each row of the modified binarized matrix is transformed to its equivalent decimal number.These decimal numbers are checked to see if it is prime in step 4.
For the proposed Bio-Cryptography, combinedly (λ, N) in order is considered as the encryption key.Alternatively, the reverse combination of encryption key, i.e., (N, λ) in order is considered as the decryption key.Since the same key is used for both encryption and decryption (but in reverse order) hence the proposed Bio-Cryptography is a symmetric key cryptographic algorithm.The key λ is divided into 128-bit length equal size subkeys, (λ 512 ..λ 385 ), (λ 384 ..λ 257 ), (λ 256 ..λ 129 ), (λ 128 ..λ 1 ) (MSB to LSB in order).Similarly, the key N (2048-bits) is divided into 128-bit length equal size subkeys, (N 2048 ..N 1921 ), (N 1920 to LSB in order).Hence, this algorithm consists of 512/128=4 subkeys from λ and 2048/128=16 subkeys from modulus N.For the proposed key generation scheme, the number of subkeys should be ≥ 2 for better encryption.Alternatively, more subkeys may increase encryption/decryption time.
From Algorithm 3, it is clear that the execution time complexity depends on the number of elements present in the vector D, which contains equivalent decimal values of the binarized matrix.Hence, line 6 of Algorithm 3 will execute n times (n = elements in D), i.e., the time complexity of line 6 is O(n).Similarly, line 7 will execute √ n times for each decimal value.Therefore, the overall time complexity to find prime numbers from n decimal values is O(n √ n) = O(n [3/2] ).b) Encryption: The proposed encryption process follows only sliding-based bit-wise XOR operation (S-XOR) recursively.This sliding operation performs from left to right direction (MSB to LSB) by the encryption keys λ and N, respectively.To perform S-XOR operation, each elements (d i ) of the feature vector Z F are transformed into 256-bit binary [b 256 , b 255 , . . ., b 1 ] in such a way that the bit lengths of the elements are more than the bit lengths of each subkeys In S-XOR encryption (Algorithm 4), the execution time of lines 3 and 12 depends on the number of enrolled image samples (i.e., constant times), which is 1-6 image samples for CVL, 1-13 image samples of FEI, and 1-4 image samples for FERET database.So, the time complexity of lines 3 and 12 is O(1).Similarly, the execution time of lines 5 Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

Algorithm 4 Encryption Using S-XOR Scheme
Input: Cancelable feature vector z F Output: Encrypted feature vector Transform each decimal d i ∈ Z F to 256-bit, encryption key λ 512-bit and modulus N to 2048-bit binary form, respectively i.e. for k ← 1 to r do /*r=(bit-length in N)/(sub-key length) 6: end for 8: end for 9: end for 10: Perform S-XOR between intermediate decrypted feature vector D and λ for step-2 decryption to obtain final decrypted feature vector.11: for i ← 1 to m do 12: for j ← 1 to n do 13: for k ← 1 to r do /*r=(bit-length in λ)/(sub-key length) 14: end for 16: end for 17: end for 18: Transform D F (=Z F ) to original cancelable features vector Z F and 13 depends on the dimension of the feature vector (i.e., variable times) which is O(n), and the execution time of lines 6 and 14 depends on the number of subkeys or the number of rounds (2, 4, 8, 16,. . ., i.e., constant times) S-XOR

V. EXPERIMENTS
In this section, we will discuss and analyze the experimental process and the outcomes of the proposed methodology in detail.The hardware platforms used were an Intel Core i5 CPU running at 3 GHz and 8-GB RAM operating at 2667 MHz.The software platforms used were MATLAB R2022a and Windows 10 Pro operating system.

A. Databases
In this work, three benchmark facial databases, namely, CVL [32], FEI [33], and FERET [34], to conduct experiments and evaluations.The CVL database consists of 114 subjects, where each subject has seven color images [Fig.6 for 994 individuals, with frontal position, profile rotation, distinct appearance, accessories, eyeglasses, varying poses, and expressions.These databases are summarized in Table I.

B. Results and Discussion
For the performance evaluation of the employed feature extraction technique, we computed both dissimilarity and similarity scores of all the databases.Table II    horizontally segmented feature vectors (f V ), and horizontally followed by vertically segmented feature vectors (f HV ).From Table II, it has been found that f V ∈ R 1×500 performs better than f 1×500 , f H ∈ R 1×500 , and f HV ∈ R 1×1000 for all the databases.Now, Table III presents the results corresponds to both CRR (%) and EER performances for different ratio of trainingtesting samples.This table also shows that the feature vector f V ∈ R [i.e., due to the vertical image segmentation (S V )] performs better than both the performances of f H ∈ R and f HV ∈ R for all the databases.Since the feature vector for further experimentation in the proposed FRS.In support of this, an experiment of comparisons has been performed corresponds to each employed database of the proposed and other existing methods in Table VII, that shows the superiority of the proposed system due to employed feature representation scheme.
Inspired by the performance reported in Table III, the feature vectors f V have been used for further process.Since the computed feature vectors f V (f F ) are the original biometric features for each person, to keep these features from The performance of CFRS 1 is quite better than that of the original feature vector f F and it is also difficult to revert f F from computed f .Hence, this computation provides quite a security to f .Table IV shows the performance obtained in CFRS 1 for all databases.
To achieve better security and improve performance, (1) has been extended to (2).In this step, we employed a permutation function π on f based on a token t 1 , i.e., π t 1 (f ) which computes f .This computation is called here as CFRS 2 .The performance of CFRS 2 is better than that of CFRS 1 , and it is also difficult to revert f from the computed f .Hence, this computation provides more security to f than f .Table V shows the performance obtained in CFRS 2 for all databases.In the further implementation of FaceHashing, (2) extended to (3).In this step, we employed a random permutation function π with another token t 2 on f , i.e., π t 2 (f ) to compute Z F .This computation is called here CFRS 3 .The generated feature vector Z F is called a cancelable features vector, and having three-tier security Z F is secured enough.Moreover, it is very difficult or almost impossible to revert f F from Z F .The extension of CFRS 2 to CFRS 3 also improves performance significantly for all the databases.Hence, the CFRS 3 system provides both 1) outstanding performance and 2) optimal security with a minimum dimension of the cancelable features vector.Both CRR (%) and ERR performance of CFRS 3 have been demonstrated in Table VI.Table VI shows that the CFRS 3 outperforms for the 200-D feature vector and sufficiently identifies the subjects with 100% accuracy.
The feature vector Z F is highly secured against reply attack, preimage attack, and record multiplicity attack.But still, there are some possibilities of several attacks, such as channel attack, database attack, etc. [49] by the intruder.Hence, we

C. Complexity Analysis of the Proposed HTPS
From (1), the time complexity of level-l FaceHashing, i.e., projection operation, is   The cancelable biometric must hold four necessary and sufficient criteria for the security, such as reusability, noninvertibility, unlinkability, and performance preservation.
1) Noninvertibility: If cancelable templates and tokens are compromised then attacker cannot revert the original biometric from the compromised template.Here, it is not possible to revert f to f or f to the original feature vector f F . 2) Reusabiltiy: A new template can be generated by assigning a new token from original biometric features kept offline for future references.3) Unlinkability: The unlinkability property says that if the same biometric is used in several biometric-based applications, there is a possibility of cross-matching over the network.The proposed system uses two unique tokens t 1 and t 2 to achieve unlinkability.4) Performance Preservation: The performance of cancelable biometrics must not be degraded compared to the original biometric.

2) Security Analysis of the Bio-Cryptographic Algorithm:
The selection criterion of cryptographic algorithm are 1) less time and space complexity; 2) high security; and 3) nature (how much confidential) and type (video, image, or text) of data.A comparison of various security issues of the ElGamal, RSA, RC5, and S-XOR algorithms are summarized in Table X.

VI. CONCLUSION
This article introduces a novel Bio-Cryptographic algorithm and key generation scheme for secure FRS in IoT environments.The major novelties of the proposed system are the implementations of 1) the S-XOR scheme as the secure, faster, and reliable Bio-Cryptographic method and 2) the user biometric-based key generation scheme.The time complexity for key generation of the proposed system is O(n √ n) to generate four prime numbers from the set of n decimal numbers.Hence, the S-XOR algorithm reduces Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

Fig. 3 .
Fig. 3. Block diagram of the proposed secure FRS in IoT environments.(a) Enrollment process.(b) Authentication process.
(a)] with four side views (angle 45 • , angle 135 • , far left, and far right) and three frontal views (smile-showing teeth, smile-showing no teeth, and serious expression) for each subject.The FEI database contains 14 color images [Fig.6(b)] of 200 individuals with varying appearances, hairstyles, and adornments captured from profile rotation up to 180 • and an upright frontal position on a homogeneous white background.The images are captured with profile rotation up to 180 • and in an upright frontal position on a homogeneous white background.The FERET face database consists of five color images [Fig.6(c)],
1: Resize image samples f F to dimension 512 × 512.2: Binarize the sample images and form a single binary matrix.3: If LSB is zero then replace it with 1 to make it odd number which increases the probability to get prime number.4: Convert each binary row to its equivalent decimal values.
presents the dissimilarity and similarity score measures of the original feature vectors (f ), vertically segmented feature vectors (f V ), Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE III PERFORMANCE
OF THE PROPOSED SYSTEM IN CRR (%) AND EER CORRESPONDS TO THE WAYS OF SEGMENTS S V , S H , AND S HV

TABLE IV PERFORMANCE
OF THE PROPOSED CFR 1 IN CRR (%) AND EER, dim STANDS FOR DIMENSION OF THE FEATURE VECTORTABLE V PERFORMANCE OF THE PROPOSED CFRS 2 IN CRR (%) AND EER, dim STANDS FOR DIMENSION OF THE FEATURE VECTOR misuse and external attacks, the proposed cancelable biometric techniques (FaceHashing) (1)-(3) have been applied to get cancelable biometric feature vectors.Here, we employed the FaceHashing technique to protect f F .To implement FaceHashing, at first, a user token (t subject )-based random matrix is generated.Then, this matrix is normalized by the Gram-Schmidth Orthogonalization method.This normalized matrix is multiplied on the columns of the original feature vector f F which computes f and this computation is called CFRS 1 .

TABLE VI PERFORMANCE
OF THE PROPOSED CFRS 3 IN CRR (%) AND EER, dim STANDS FOR DIMENSIONAL OF THE FEATURE VECTOR Table VIII summarized the time complexities of ElGamal, RC5, RSA, and S-XOR algorithms.From Table VIII, it is clear that the proposed Bio-Cryptography's time complexity is much less compared to other competing methods.For better understanding, Table IX demonstrates these comparisons in terms of Second corresponds to each employed database.

TABLE IX EXECUTION
TIME (IN SECONDS) COMPARISON BETWEEN PROPOSED AND EXISTING CRYPTOGRAPHIC ALGORITHMS D. Security Analysis of the Proposed HTPS 1) Security Analysis of the Cancelable Biometrics:

TABLE X COMPARATIVE
STUDY OF THE PROPOSED ENCRYPTION METHOD AND ELGAMAL, RC5, AND RSA ALGORITHMS.INDICATES ATTACK POSSIBLE AND × INDICATES ATTACK IS NOT POSSIBLE the overall encryption/decryption time.From the observations in Table II, the original feature vector provides 97.40%, 96.91%, and 98.27% accuracy for the CVL, FEI, and FERET databases, respectively.According to the performance reported in Tables IV and V, CFRS 1 provides 98.73%, 98.01%, and 99.84% accuracy, and CFRS 2 provides 99.47%, 98.10%, and 100% accuracy for CVL, FEI, and FERET databases, respectively.Whereas, according to the performance reported in Table VI, CFRS 3 provides 100% accuracy for all databases regarding the 200-D feature vector.The proposed FRS outperforms after the implementation of CFRS 3 and CFRS 3 is more secure than f V .In future work, an efficient deep-learning-based cancelable face recognition approach can be implemented with the enhancement of the security of the FRS in IoT environments.