Cancelable biometrics are required in most remote access applications that need an authentication stage such as the cloud and Internet of Things (IoT) networks. The objective of using cancelable biometrics is to save the original ones from hacking attempts. A generalized algorithm to generate cancelable templates that is applicable on both single and multiple biometrics is proposed in this paper to be considered for cloud and IoT applications. The original biometric is blurred with two co-prime operators. Hence, it can be recovered as the Greatest Common Divisor (GCD) between its two blurred versions. Minimal changes if induced in the biometric image prior to processing with co-prime operators prevents the recovery of the original biometric image through a GCD operation. Hence, the ability to change cancelable templates is guaranteed, since the owner of the biometric can pre-determine and manage the minimal change induced in the biometric image. Furthermore, we test the utility of the proposed algorithm in the single- and multi-biometric scenarios. The multi-biometric scenario depends on compressing face, fingerprint, iris, and palm print images, simultaneously, to generate the cancelable templates. Evaluation metrics such as Equal Error Rate (EER) and Area and Receiver Operator Characteristic curve (AROC) are considered. Simulation results on single- and multi-biometric scenarios show high AROC values up to 99.59%, and low EER values down to 0.04%.

Compared to authentication systems based on passwords, tokens, IDs, and biometrics, cancelable biometric recognition systems provide better security for human identification purposes. They are more suitable for remote access networks such as cloud and IoT networks. Cancelable biometric systems can be built with a single or uni-biometric, hence the name uCBS, or with multiple biometrics, hence the name mCBS. The uCBS are considered less secure compared to the mCBS [

User validation schemes are impeded by overlapping facial biometrics as in twins, poor data acquisition in the case of dry fingerprints, or even missing data, in typical cases of occluded biometric images. Therefore, mCBS are required to enhance the outcomes of the recognition process. Human facial, iris, fingerprint, ear, signature, voice, and other biometric modalities are now widely exploited to build robust mCBS. At the same time, such behavioral biometrics suffer from unavailability or poor coverage over large databases, and they also exhibit poor recognition accuracy [

For increasing the security, accuracy, and genuine acceptance rate, mCBS have been implemented [

Furthermore, Yang et al. [

Similarly, Lai et al. [

This paper presents a new algorithm that can be applied for both uCBS and mCBS. The main idea of the proposed algorithm depends on the Greatest Common Divisor (GCD) to generate the cancelable templates. It is known that if two co-prime operators are used on the same biometric image to obtain two blurred versions of that image, the original biometric itself can be obtained again through the 2D GCD between the two blurred versions. If some intended change is induced in the biometric image prior to or after the application of one of the blurring operators, this will lead to a distorted output from the 2D GCD operation. This output can be used instead of the original biometric template as a cancelable template. This idea is adopted in this paper to build uCBS and mCBS. The remainder of this study is outlined as follows. The basics of the related 1-D and 2-D Sylvester GCD algorithms are discussed in Section 2. The proposed cancelable template generation algorithm for uCBS and mCBS is presented in Section 3. Extensive simulation experiments are presented in Section 4 to validate the proposed algorithm. Finally, the concluding remarks are summarized in Section 5.

This section introduces the major fundamental theories of the 1-D and 2-D Sylvester GCD algorithms that will be exploited in the proposed cancelable biometric algorithm to create the distorted versions of the original biometrics.

Let

and

Similarly, if the GCD between the two polynomials is equal to

where

and

are two relatively co-prime polynomials. Consequently, it follows from [

By equating the coefficients of like powers of

where

Hence,

By analyzing the matrix

In some cases, the degree

where

with

Therefore,

Using arguments in

The direct extension of the 1-D Sylvester algorithm to the 2-D case in terms of constant matrices generated from the given 2-D polynomial coefficients leads to very large-size matrices. For

In doing so, we assume that the two blurred versions

where

It is trivial that _{1}

We scale each row by a constant

Undertaking similar operations and substituting

From

where

and

Multiplying

For an eigenvector

Finally, the inverse Fourier transform can be used to estimate the GCD. If we use two blurred images of the same biometric and estimate their 2D GCD, we can get the original biometric template. On the other hand, if we make a slight change induced by the user prior to or after blurring, we severely distort the 2D GCD result. The result of the GCD in this case can be used as a cancelable template. The minor changes can be induced in each biometric template in a user-specific manner.

Applying

where

and

If

and

As shown in

In this section, the assessment of the suggested algorithm is introduced. Firstly, the security analysis of the suggested GCD-based algorithm as an encryption-like algorithm is presented in terms of visual analysis, correlation analysis, differential attack analysis, and entropy analysis [

Furthermore, in this section, several experiments are introduced to verify the validity of the proposed uCBS and mCBS that depend on 2D GCD. They have been implemented using a workstation equipped with MATLAB Intel ® Core ™ i7-4210U on a CPU with a 1.7 GHz processor. Four datasets have been used in the uCBS experiments, namely ORL [

For the uCBS, the experiments are based on generating two blurred versions of each template and inducing a minor change in one of them prior to blurring. Hence, the 2D-GCD is implemented to generate the cancelable template of that biometric. The database of cancelable templates is composed, and hence the distance between new templates and those in the database is estimated based on the correlation score. Both EER and AROC values are estimated for the verification process. On the other hand, the proposed mCBS depends on estimating the DCTs of four biometric templates and generating a combined version based on the first quadrant of each DCT. This combined version is used as an initial template that is blurred twice. A minor change in one of these versions and the application of the 2D-GCD lead to the cancelable template.

The Receiver Operating Characteristic (ROC) curve, which represents the relationship between the true-positive correlation and false-positive correlation [

The multi-modal biometrics used to validate the proposed CBS consist of facial, fingerprint, iris, and palm print images as presented in

Finally, the average A ROC, mean correlation scores, False Acceptance Rate (FAR), False Rejection Rate (FRR), and ERR for the GCD-based mCBS are presented in

System | AROC | Mean of authorized correlation score | Mean of un-authorized correlation score | FAR | FRR | ERR |
---|---|---|---|---|---|---|

Proposed uCBS for face images | 0.9806 | 0.8278 | 0.0302 | 0.0916 | 0.0328 | 0.0098 |

Proposed uCBS for fingerprint images | 0.9922 | 0.8799 | 0.0417 | 0.0401 | 0.0107 | 0.0071 |

Proposed uCBS for iris images | 0.9943 | 0.8817 | 0.0430 | 0.0476 | 0.0230 | 0.0107 |

Proposed uCBS for palm print images | 0.9780 | 0.8186 | 0.0225 | 0.0735 | 0.0129 | 0.0096 |

Proposed mCBS | 0.9802 | 0.8529 | 0.0031 | 0.0933 | 0.0092 | 0.0049 |

To validate the suggested CBS, more test investigations have been carried out for comparison of the suggested CBS with the latest algorithms [

CBAS | EER | FAR | FRR | AROC |
---|---|---|---|---|

Proposed | 0.0023 | 0.0182 | 0.0024 | 0.968 |

[ |
0.0046 | 0.0235 | 0.0929 | 0.883 |

[ |
0.0357 | 0.0612 | 0.863 | |

[ |
0.0859 | 0.0627 | 0.718 | |

[ |
0.0416 | 0.1955 | 0.0489 | 0.873 |

[ |
0.1081 | 0.0927 | 0.0967 | 0.907 |

[ |
0.0924 | 0.0562 | 0.0257 | 0.868 |

[ |
0.0178 | 0.0571 | 0.0876 | 0.896 |

[ |
0.0098 | 0.0104 | 0.018 | 0.952 |

This paper presented a new approach to build efficient CBS using single- and multi-biometric inputs for cloud and IoT biometric applications. Pre-determined distortions are induced in the biometric images for single- and multi-biometric inputs with the GCD algorithm. As a self-dependent approach, the need for auxiliary data or images is eliminated. The GCD with some minimal changes can be used efficiently in the generation of cancelable biometric templates. We validated the proposed uCBS and mCBS on inputs consisting of facial, fingerprint, iris, and palm print images. AROC values above 99% were recorded for all the examined biometrics. This work can be easily implemented for cloud, IoT, and wireless access applications. In addition, it can be enhanced with the utilization of encryption algorithms with the GCD algorithm. In the future, we can incorporate deep learning algorithms for compressing and encrypting the biometric images for enhancing the cancelable biometric system performance.

The authors would like to thank the support of the Deanship of Scientific Research at Princess Nourah bint Abdulrahman University.