Generation of Privacy-friendly Datasets of Latent Fingerprint Images

using Generative Adversarial Networks

Stefan Seidlitz, Kris Jürgens, Andrey Makrushin, Christian Kraetzer and Jana Dittmann

Otto-von-Guericke University Magdeburg, Universitaetsplatz 2, Magdeburg, Germany

Keywords: Digitized Forensics, Latent Fingerprint, Image Synthesis, Generative Adversarial Networks, GAN, Privacy.

Abstract: The restrictions posed by the recent trans-border regulations to the usage of biometric data force researchers

in the fields of digitized forensics and biometrics to use synthetic data for development and evaluation of new

algorithms. For digitized forensics, we introduce a technique for conversion of privacy-sensitive datasets of

real latent fingerprints to "privacy-friendly" datasets of synthesized fingerprints. Privacy-friendly means in

our context that the generated fingerprint images cannot be linked to a particular person who provided

fingerprints to the original dataset. In contrast to the standard fingerprint generation approach that makes use

of mathematical modeling for drawing ridge-line patterns, we propose applying a data-driven approach

making use of generative adversarial neural networks (GAN). In our synthesis experiments the performance

of three established GAN architectures is examined. The NIST Special Database 27 is exemplary used as a

data source of real latent fingerprints. The set of training images is augmented by applying filters from the

StirTrace benchmarking tool. The suitability of the generated fingerprint images is checked with the NIST

fingerprint image quality tool (NFIQ2). The unlinkability to any original fingerprint is established by

evaluating outcomes of the NIST fingerprint matching tool.

1 INTRODUCTION

Fingerprints are known to be directly linked to an

individual and fingerprint as biometric modality is

well accepted and widely established means of user

authentication. The applications making use of

fingerprints spread form biometric access control

systems to forensic investigation of latent

fingerprints. Empirical studies on forensic or

biometric fingerprint processing and recognition

require a large dataset of fingerprints for validation of

results. Moreover, development of fingerprint

detection and recognition algorithms based on

machine learning is hardly possible without an

abundant amount of training data. However,

fingerprints as well as any other biometric data are

seen as a special category of personal data which is

prohibited to be processed by the recent trans-border

regulations without exception for the purpose of non-

commercial research. A prominent example of a

fingerprint dataset valuable for digitized forensics

which was removed from the public access after

establishing such regulations is the NIST Special

Database 27 (Garris et al., 2000). Note that some

categories of personal data can be processed after

anonymization which is not the case for biometric

samples because they require no meta-data to be

linked to individuals. An elegant solution to the

privacy-caused processing restrictions is a genera-

tion of artificial fingerprints that have the same

characteristics as real fingerprints, but cannot be

linked to particular persons. For biometrics, there is a

de facto standard synthesizing tool called SFinGe

(Cappelli, 2009). Fingerprints are generated to fit a

certain basic pattern and a predefined set of minutitae.

In contrast, in the field of digitized forensics the focus

is on mimicking substrate and environmental

influences on a digitalized latent fingerprint. Since

such influences can be hardly formalized, modern

data-driven image generation approaches has to be

adopted for fingerprints.

Recent achievements in development of deep

convolutional neural networks and especially gene-

rative adversarial networks (GAN) allow for

automated generation of artificial images that can be

hardly told apart from real images. In this paper, we

examine several GAN architectures in application to

generation of fingerprint images and assess the

quality of the generated images. The quality is two-

fold. The generated fingerprints must appear natural

Seidlitz, S., Jürgens, K., Makrushin, A., Kraetzer, C. and Dittmann, J.

Generation of Privacy-friendly Datasets of Latent Fingerprint Images using Generative Adversarial Networks.

DOI: 10.5220/0010251603450352

In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2021) - Volume 4: VISAPP, pages

345-352

ISBN: 978-989-758-488-6

345

and be privacy-friendly. Natural appearance means

that the naked eye cannot see difference to a typical

original pattern and a fingerprint is applicable for the

further investigations (appropriate basic pattern,

sufficient number of minutiae, etc). Privacy-friendly

means that a generated fingerprint does not match one

particular original fingerprint. If a generated

fingerprint matches k (k ≠ 1) original instances, then

k-anonymity will be indicated. As a metric for both

criteria we use NIST tools (Ko, 2007): NFIQ2 for the

former and the combination of the MINDTCT and

Bozorth3 for the latter.

Our contribution is in demonstrating that GAN

can be successfully applied for conversion of privacy-

sensitive datasets of fingerprint images to privacy-

friendly datasets and in comparing three currently

very prominent GAN architectures:

ProgressiveGAN, StyleGAN and StyleGAN2 for this

purpose. While comparing generation perfor-mances

of the networks, we consider the following

characteristics: how many training iterations are

required to obtain high-quality fingerprint images, the

time of image generation, and the proportion of high-

quality fingerprints in the whole number of generated

fingerprints. The better network would be able to

generate more high-quality fingerprints in a shorter

time frame. In order to improve the diversity of

generated images and make the process of image

generation more stable, we augment the training set

by applying filters from the StirTrace benchmarking

tool (Hildebrandt et al., 2015). Note that here we are

not focused on generation of one particular dataset of

artificial fingerprints, but rather propose a technique

for compilation of such privacy-friendly datasets out

of existing data.

In Section 2, we overview related works. In

Section 3, we introduce our concept of generation and

assessment of synthetic fingerprint images. In Section

4, we elaborate on important aspects of our

implementation. In Section 5, we evaluate the GAN

generated fingerprint images. Section 6 concludes the

paper with the summary of results.

2 RELATED WORKS

Early works on generation of synthetic fingerprint

images were concerned rather with reconstruction of

ridge-line patterns from biometric templates

considering this act as a potential attack on a

biometric system (Galbally et al., 2008). Starting

from the set of minutiae, a fingerprint area, an

orientation map and a frequency map are estimated.

Then, the iterative pattern growing approach draws

ridges along the orientation lines by applying Gabor

filters (Cappelli et al., 2007). This approach has been

implemented in the software called SFinGe (Cappelli,

2009). The most critical step in this workflow is the

estimation of an orientation map based on the set of

minutiae or, to be more precise, based on singular

points of a basic pattern (core, deltas). It is shown in

(Ram et al., 2010) that singular points can be modeled

by the zero-poles of Legendre polynomials, resulting

in a discontinuous orientation field. An exhaustive

study on the possibility of modeling fingerprints by

the phase portraits of differential equations is

conducted in (Zinoun, 2018) and the limitations are

outlined.

In contrast to mathematical modeling, the modern

trend is a data-driven generation of realistic images

by means of generative adversarial networks (GAN).

Recently, very impressive results have been

demonstrated with images of human faces. Based on

a huge amount of face images, researchers from

NVIDIA successfully created high-quality, high-

resolution synthetic faces, which can be hardly told

apart from the real ones (Karras et al., 2018). The first

effort to synthesize fingerprints using a Wasserstein

GAN is made in (Bontrager et al, 2017) aiming at

generating so-called master fingerprints that match

multiple original fingerprints. Later on, in (Minaee et

al., 2018), a connectivity imposed GAN is introduced

and applied to two datasets: FVC-2006 and PolyU.

The size of generated images in both publications is

rather insufficient. In (Attia et al, 2019) fingerprints

are synthesized by a variational autoencoder. In (Cao

et al., 2018), a combination of an autoencoder and an

adapted Wasserstein GAN is used for synthesizing

512x512 pixel fingerprint images. A CycleGAN is

applied in (Wyzykowski et al., 2020) to transfer

texture from real fingerprints to conventionally

synthesized ridge-line patterns with added sweet

pores which dramatically improves their realistic

appearance. An alternative approach for generating

high-resolution realistic fingerprints is in combining

GAN with a super-resolution network proposed in

(Riazi et al. 2020). In (Fahim, et al., 2020), a

lightweight GAN is proposed for creating 128x128

pixel images and compared with five established

GAN architectures based on 64x64 pixel patches. The

next breakthrough is done in (Mistry et al., 2020) by

incorporating identity information into the fingerprint

synthesis network which is based once again on

combining auto-encoder and Wasserstein GAN.

Here, we look for a suitable GAN architecture to

generate high-resolution (512x512 pixel) gray-scale

fingerprint images. This target size corresponds to the

high-quality fingerprint image format (1000 ppi) used

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

346

in dactyloscopic analyses (Orandi et al., 2014).

Images with this size and resolution are capable of

representing a significant part of a scanned latent

fingerprint incl. the level 1 feature (basic pattern) as

well as a sufficient number of level 2 features

(minutiae) and potentially, depending on the prints

quality, also level 3 features (sweat pores). The vast

majority of dated GAN architectures are designed to

generate images with a resolution less or equal to

256x256 pixels (see e.g. (Karras et al., 2017)) and

therefore omitted in our considerations. Generating

images of higher resolution is possible by adding

several up-sampling layers (Zhang et al., 2016), but it

would increase instability of the training process.

Some GAN architectures such as ConditionalGAN

(Wang et al., 2018) suffice the resolution criterion but

are not capable of generating plausible images from

fully random latent vectors. Such GANs are also not

addressed. To the best of our knowledge, currently

only three GAN architectures fit to our requirements:

ProgressiveGAN (Karras et al., 2017), StyleGAN

(Karras et al., 2018) and StyleGAN2 (Karras et al.,

2019). These networks are able to generate plausible

high-resolution images from fully random latent

vectors.

3 CONCEPT

The usage of GAN is driven by privacy concerns. The

synthesized fingerprints in a privacy-friendly dataset

should exactly reproduce the characteristics of a

reference dataset, but must leak no information in

terms of reproducing the same minutiae.

The classic fingerprint synthesizing approaches

start with minutiae or singular points to generate

ridge-lines. If a set of minutia is taken from a

reference fingerprint, the synthesized fingerprint

would perfectly match it. Random selection of

singular points or of a set of minutiae may lead to

generation of implausible ridge-line patterns. Hence,

tools like SFinGe can be successfully applied for

generation of fingerprints with parameterizable

characteristics reaching the diversity by randomizing

parameters. However, SFinGe is hardly applicable to

mimic characteristics that are presented in some

reference data but cannot be formally described such

as an appearance of a substrate. This is often a case

for latent fingerprints which include not only

characteristic of the fingerprint itself, but also

characteristics of the environment in which the

fingerprints were left behind (including topological

characteristics of the surface). Table 1 summarizes

the differences between the both aforementioned

concepts of synthesizing fingerprint images.

Table 1: Comparison of SFinGe and GAN for generation of

fingerprint images.

SFinGe GAN

Generation

approach

mathematical

modeling

data-driven

Fingerprint type exemplar latent

Reproduction of characteristics of

a fingerprint

(basic pattern,

minutiae)

characteristics of

environment

(incl. substrate,

digitalization

process etc.)

Our concept for the generation of privacy-friendly

fingerprint image datasets is illustrated in Figures 1

and 2. The basic idea is that, for a given proprietary

dataset of fingerprint images, we create a dataset of

anonymous fingerprint images that are not linked to

individuals but preserve all characteristics of the

images in the initial dataset incl. background noise,

image quality, frequency of ridge lines, one of the

standard basic patterns, plausible number and

locations of minutiae, etc. The number of samples in

a new dataset is optional and depends only on time

spent to the generation process. Note that images in

the original dataset may include more than one

fingerprint and may also vary in size. In contrast, the

GAN training images must all have a specific size.

Hence, a dataset specific pre-processing of images is

required.

Figure 1: Training of a GAN model.

Figure 1 schematically describes training of a

GAN model. In our further considerations, we take

NIST Special Database (SD) 27 (Garris, 2000) with

2856 latent and matching tenprint fingerprint images

as an example source of fingerprint data. There are

two pre-processing steps performed before images

are fed into the network: First, the fingerprints are

segmented from the images. The patches are cut

around fingerprint core points so that each resulting

image contains exactly one partial or full fingerprint.

The core points are located using the NIST tool

MINDTCT (Ko, 2010). Note that the MINDTCT

sometimes falsely highlight some artifacts like letters

NIST SD 27,

1000 dpi

(2856 images)

FP patches,

1000 dpi,

512×512 pixels

(39768 samples)

GAN Model

FP patches

1000 dpi,

512×512 pixels

(516984 samples)

Segmentation Augmentation

(StirTrace)

Training

of GAN

NIST SD 27,

1000 dpi

(2856 images)

FP patches,

1000 dpi,

512×512 pixels

(39768 samples)

GAN Model

FP patches

1000 dpi,

512×512 pixels

(516984 samples)

Segmentation Augmentation

(StirTrace)

Training

of GAN

Generation of Privacy-friendly Datasets of Latent Fingerprint Images using Generative Adversarial Networks

347

on the scanned trace cards in NIST SD 27 images as

core points leading to non-fingerprint patches.

However, we decided not to remove these outliers

from the training data because they should be

automatically sorted out in the later steps of the

generation process. The number of extracted

fingerprint patches is 39768. Second, the resulting set

of fingerprint patches is augmented by their filtered

versions (see Section 3.1) increasing the number from

39768 to 516984 for 12 StirTrace filters applied. This

is done to improve the diversity of the generated

images and to make the process of image generation

more stable.

Figure 2: Generation of synthetic fingerprint images.

Figure 2 demonstrates the generation process of

fingerprints and assessment of their quality. During

GAN training, random latent vectors are fed into the

generator resulting in synthetic fingerprint images.

The snapshots of generated fingerprints are automa-

tically stored at some training iterations. From each

snapshot, we randomly pick 1000 fingerprint images

for further analysis. For every image in every

snapshot, we obtain a quality score using the NIST

Fingerprint Image Quality estimator - NFIQ2 (Elham

et al., 2013). From each snapshot, we calculate the

mean value of the NFIQ2 scores and the number of

images with NFIQ2 scores higher than 35 to

determine at which iteration of the training process

the generation model works best. Images with such

scores correspond to the two highest quality classes

in the 5-class NFIQ scale (Galbally et al., 2019) and,

therefore, are referred to as high-quality fingerprints.

The best snapshot has the highest number of high-

quality 512x512 pixel fingerprints. Fingerprints with

NFIQ2 scores lower or equal to 35 are filtered out.

Each remaining fingerprint is biometrically matched

to all training fingerprints using NIST tools (see

Section 3.3).

3.1 Data Augmentation with StirTrace

StirTrace (Hildebrandt et al., 2015) is designed for

benchmarking pattern recognition tasks in the context

of digitized forensics. This tool can be seen as a set of

filters to mimic typical artifacts that usually arise in

the process of digitizing fingerprints, especially

addition of noise. The most relevant filters used here

for data augmentation are additive noise (strengths: 3,

5 and 9), additive Gaussian noise (strengths: 3, 5 and

9), median cut (strengths 3, 5 and 9) and salt and

pepper noise (strengths: 3, 5 and 9). Note that filtering

does not change the number and location of minutiae

in a fingerprint. Hence, it is sufficient that the

generated synthetic fingerprints are matched only

against original patches and not against all training

samples.

3.2 GAN Architectures

The term Generative Adversarial Network (GAN)

was proposed in 2014 by Ian Godfellow for the

architecture containing two neural networks

Generator (G) and Discriminator (D), which are

trained interchangeably. Generator produces synthe-

tic data from random vectors and discriminator tries

to distinguish these data from genuine data. The basic

idea is that the generator improves with training while

the discriminator’s performance gets worse.

However, after the point where the discrimi-nator is

unable to tell apart genuine and synthesized data, the

generator cannot be improved further, making the

training process rather unstable. When operating with

images, D is represented by a convolutional neural

network and G by a de-convolutional neural network.

Progressive growing GAN (ProgressiveGAN) is

an elegant solution to the convergence issue of the

GAN training (Karras et al., 2017). Training begins

with low-resolution images e.g. 4x4 pixels and the

input images are re-scaled to this resolution. After the

training process converges at the selected resolution,

the resolution is increased and the training is repeated.

This is done until the target resolution is reached. This

is how the generator learns rough characteristics of

training images first and then gradually fine

characteristics. Technically, switching of resolution

happens by gradual addition of intermediate layers

into the network. Progressive-GAN was the first

architecture that enabled gene-ration of high-

GAN

Generator

Random

latent

vectors

Snapshots

(FP patches,

1000 dpi)

Avg. NFIQ2 scores

for snapshots

Snapshot X

(FP patches,

1000 dpi)

FP patches,

1000 dpi

(39768 samples)

….

Snapshot X

(FP patches,

1000 dpi)

FP patches,

500 dpi

(39768 samples)

Snapshot X

(FP patches,

500 dpi)

Calculation

of NFIQ2

scores

Snapshot

selection

Filtering

(NFIQ2 > 35)

Down-

sampling

Down-

sampling

MINDTCT

Minutiae

(39768 samples)

Minutiae

(1000 samples)

Calculation

of Bozorth3

scores

Bozorth3

scores

….

High-quality,

anonymous FP,

1000 dpi

Filtering

(No matches)

GAN

Generator

Random

latent

vectors

Snapshots

(FP patches,

1000 dpi)

Avg. NFIQ2 scores

for snapshots

Snapshot X

(FP patches,

1000 dpi)

FP patches,

1000 dpi

(39768 samples)

….

Snapshot X

(FP patches,

1000 dpi)

FP patches,

500 dpi

(39768 samples)

Snapshot X

(FP patches,

500 dpi)

Calculation

of NFIQ2

scores

Snapshot

selection

Filtering

(NFIQ2 > 35)

Down-

sampling

Down-

sampling

MINDTCT

Minutiae

(39768 samples)

Minutiae

(1000 samples)

Calculation

of Bozorth3

scores

Bozorth3

scores

….

High-quality,

anonymous FP,

1000 dpi

Filtering

(No matches)

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

348

resolution naturally looking fake faces which can be

hardly told apart from real ones.

Based on ProgressiveGAN, an improved archite-

cture called StyleGAN (Karras et al., 2018) was

developed to take apart aggregated characteristics of

images also referred to as styles and to move from one

style to another. For face images the styles are e.g. a

haircut or a skin color. Technically, the analysis and

clustering of the latent space is done by introduction

of adaptive instance normalization (AdaIN) layers

and addition of noises to control the intensity of a

particular style. However, StyleGAN often produces

characteristic imperfection e.g. droplet or phase

artifacts which are attributed to the network

architecture. Droplet artifacts arise due to

independent normalization of means and variations of

different style feature maps in AdaIN layers and

phase artifacts arise due to progressive growing. A

variation of the StyleGAN architecture called Style-

GAN2 (Karras et al., 2019) was proposed to avoid the

aforementioned imperfections. Adaptive instance

normalization is replaced by weight demodulation

making normalization of means and variations of the

different styles not independent anymore. The

progressive growing issue is solved by generating

images with only target resolution by adding up the

weighted outcomes of all layers of the generator. It

does not change the idea of progressive learning of

image characteristics (from rough to fine), but helps

to get rid of phase artifacts.

3.3 Fingerprint Anonymity Assessment

Biometric matching of fingerprints is done by using

NIST tools: MINDTCT and Bozorth3 (Ko, 2007).

MINDTCT extracts the list of minutia from a finger-

print while Bozorth3 compares two such lists and

produces a similarity score representing the number

of matched minutiae. Note that MINDTCT requires

500 dpi images to properly detect minutiae. Since the

target resolution of our synthesized fingerprints is

1000 dpi, the images are downscaled with a factor 2

before applying MINDTCT. The documentation of

Bozorth3 suggests values over 40 for the perfect

match. However, we use here a value of 30 as a

decision threshold to guarantee better anonymity.

4 IMPLEMENTATION

We use the original NVIDIA implementations of the

addressed GAN architectures from GitHub reposito-

ries: http://github.com/NVlabs/stylegan and http://

github.com/NVlabs/stylegan2. The implementation

of ProgressiveGAN is a part of the StyleGAN repo-

sitory. The GANs are used in their default configu-

ration, we only set the target image size to 512x512

pixel and kImages to 7500. The parameter kImages

refers to the amount of real images which the GAN

discriminator has seen during training. We switch off

the estimation of the perceptual path length and linear

separability of all GANs because these metrics do not

work with gray scale images (for reasons see (Karras

et al., 2018)). The amounts of layers used and the

trainable parameters of each GAN are summarized in

Table 2. For the training of the GANs we used a

workstation with two NVIDIA Titan RTX graphic

cards with 24 GB VRAM each.

The StirTrace tool is used in version 4 as provided

at https://sourceforge.net/projects/stirtrace. The

applied filters as well as the composition of the

training set after this data augmentation step are

summarized in Table 3.

The NFIQ2, MINDTCT and Bozorth3 are the

parts of the NIST Biometric Image Software (NBIS)

which is available at https://www.nist.gov/services-

resources/software/nistbiometric-image-software-

nbis.

Table 2: Configurations of the addressed GANs.

GAN type network type amount

of layers

trainable

parameters

Progressive

GAN

Generator

Discriminator

23.067.048

23.075.169

StyleGAN Generator

Discriminator

26.174.696

23.075.169

StyleGAN2 Generator

Discriminator

30.270.551

28.982.721

Table 3: Composition of the training dataset.

Filter Kernel size Number of samples

without filter - 39768

+ additive noise 3, 5, 9 3 x 39768 = 119304

+ additive Gaussian

noise

3, 5, 9 3 x 39768 = 119304

+ median cut 3, 5, 9 3 x 39768 = 119304

Salt & pepper noise 3, 5, 9 3 x 39768 = 119304

Total: 516984

5 EVALUATION

We compare GAN architectures regarding the

following aspects: how many training iterations are

required to obtain high-quality fingerprint images, the

speed of image generation, and the proportion of

Generation of Privacy-friendly Datasets of Latent Fingerprint Images using Generative Adversarial Networks

349

high-quality fingerprints in the whole number of

generated fingerprints. The better network is able to

generate more high-quality fingerprint images in a

shorter time frame. Additionally, we validate the

“anonymity” of the created datasets.

5.1 Training and Generation Time

The training time strongly varies from one GAN to

another. Due to the time constraints, we initially

limited the training parameter kImages to 7500.

Nonetheless, only StyleGAN reached the image size

of 512x512 pixels with 7500 kImages on one GPU.

The training of ProgressiveGAN with 7500 kImages

resulted in images of 256x256 pixels. Hence, the

training was continued with 12000 kImages on two

GPUs. For the proper comparison the 256x256

images are upscaled to 512x512 pixels. The training

time of StyleGAN2 is extremely long (see Table 4),

so that we first switched from one GPU to two GPUs

after 3429 kImages and stopped training of after 4452

kImages (ca. 11 days). Nevertheless, even at this

point, the generated fingerprint images already

reached a good subjective quality (to be attributed to

the residual structure of generator and discriminator).

Note that StyleGAN2 has an advantage over the two

other architectures because training already starts

with 512x512 pixel images while the other two start

with very small images und gradually upscale them.

Table 4: Training time of the GANs on our reference PC.

GAN type Used

GPUs

kImages Training

time

Reached

image size

Progressive

GAN

1 < 7500 2d 1h 44m 256x256

Progressive

GAN

2 7501-

12000

4d 22h 57m 512x512

StyleGAN 1 < 7500 3d 3h 36m 512x512

StyleGAN2 1 < 3429 8d 5h 12m 512x512

StyleGAN2 2 3430-

4452

2d 20h 46m 512x512

5.2 Quality of Generated Images

Figures 3 and 4 demonstrate the development of the

average NFIQ2 score and the ratio of high-quality

images of 1000 images randomly selected from GAN

snapshots over the course of training, respectively.

We see that StyleGAN2 starts genera-ting high-

quality fingerprints after only few training iterations

and the fingerprint quality even degrades with the

higher kImages parameter. The diagrams also

demonstrate that StyleGAN clearly outperforms

ProgressiveGAN regarding both the average NFIQ2

score and the number of fingerprints with NFIQ2

scores higher than 35. For StyleGAN and Progres-

siveGAN, the snapshot at which the targeted

resolution of 512x512 pixels is reached is marked.

For all three GANs, the average NFIQ2 scores

stabilize between 25 and 30.

Figure 3: GAN training: average NFIQ2 scores.

Figure 4: GAN training: ratio of high-quality fingerprints.

Figure 5: Distributions of NFIQ2 scores of 1000 random

GAN fingerprint images from the selected snapshots.

We compare the performances of GANs by

comparing the snapshots at which GAN reaches the

highest average NFIQ2 score with the possibly low

kImages parameter. The optimal value of kImages is

7870, 6840 and 1925 for ProgressiveGAN, Style-

GAN and StyleGAN2 respectively. The distributions

of NFIQ2 scores for the selected snapshots are

depicted in Figure 5. The histograms shows that

512x512

6630

7840

512x512

6630

7840

512x512

6630

7840

512x512

6630

7840

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

350

fingerprint images with the highest quality are

generated by StyleGAN followed by Progressive

GAN and StyleGAN2. Out of 1000 images in a

corresponding snapshot, StyleGAN generated 371

images with NFIQ2 higher than 35, Progressive GAN

266 images and StyleGAN2 191 images. Note that all

GAN generated images have on average higher

NFIQ2 scores than original training images.

Table 5 visualizes the difference between GAN-

generated fingerprint images of high (NFIQ2 > 35)

and of relatively low quality (25 ≤ NFIQ2 < 35). The

perceptual quality of the fingerprint images is

extremely high, with plausible level 1 and level 2

features and, in the case of StyleGAN, even

something that already looks like some sweat pores.

Table 5: Examples of GAN-generated fingerprint images.

Progres.GAN StyleGAN StyleGAN2

NFIQ2

> 35

39 43

NFIQ2

< 35

26 26 25

5.3 Anonymity of Generated

Fingerprints

After filtering out the low-quality fingerprint images

with NFIQ2 score below 35, we compare each high-

quality fingerprint image with all training images.

Low Bozorth3 scores indicate that the generated

fingerprints cannot be linked to any person who

provided fingerprints to the training dataset. The

average Bozorth3 scores range between two and six

depending on the fingerprint and the GAN type.

However, there are few generated fingerprints with

exactly one Bozorth3 score higher than 30 indicating

the match between a generated fingerprint and one of

the fingerprints in the training dataset. Table 6 shows

the ratio of synthesized fingerprints with Bozorth3

scores lower than 30, between 30 and 39, and higher

than 40 for each considered GAN type. For

StyleGAN for instance, 59 images have a score above

40 meaning 15.9% of high-quality non-anonymous

images in the selected snapshot.

Table 6: Number of GAN fingerprints with Bozorth3 scores

(s) in a certain range in the selected snapshots.

Progres.GA

StyleGAN StyleGAN2

s < 30 169/266 ~

63.53%

182/371 ~

49.06%

106/191 ~

55.50%

30 ≤ s < 40 65/266 ~

24.44%

130/371 ~

35.04%

66/191 ~

34.55%

s ≥ 40 32/266 ~

12.03%

59/371 ~

15.90%

19/191 ~

9.95%

The experimental results suggest that Progressive

GAN with 63.53% of Bozorth3 scores lower than 30

generates on average the highest number of anony-

mous fingerprints and therefore can be seen as the

most privacy-friendly generation approach. Progres-

siveGAN is followed by StyleGAN2 (55.5% anony-

mous fingerprints) and then StyleGAN (49.06%

anonymous fingerprints). Considering the absolute

number of the high-quality anonymous fingerprints in

the selected snapshot, StyleGAN has clearly the best

generation performance with 182 images followed by

ProgressiveGAN (169 images) and then StyleGAN2

(106 images).

In our case study, we conducted experiments only

with the NIST SD 27 database. The proportions of

anonymous high-quality fingerprints within the

whole number of generated fingerprints as well as the

generation time cannot be generalized for any

reference database taken as training data for a GAN.

However, the experimental results clearly show that

GAN is a suitable technique for "anonymization" of

privacy-sensitive fingerprint datasets.

6 CONCLUSION

We demonstrate that a GAN is in general a suitable

technique for generation of high-quality anonymous

fingerprint images. As a data-driven approach a GAN

takes a privacy-sensitive dataset and converts it to

privacy-friendly dataset without loss of dataset

characteristics. The resulting datasets can be used for

research on fingerprints without privacy-caused

limitations. In a case study with the NIST SD27

dataset, we show that all three addressed GAN

architectures (ProgressiveGAN, StyleGAN and Style

GAN2) are capable of converting original privacy-

sensitive fingerprint images to privacy-friendly ones.

StyleGAN2 has an advantage that the fingerprint

images with high NFIQ2 scores are generated after

only a few iterations of training. In contrast,

ProgressiveGAN and StyleGAN require many

training iterations to reach the target image

Generation of Privacy-friendly Datasets of Latent Fingerprint Images using Generative Adversarial Networks

351

resolution. However, StyleGAN2 is the worst

approach regarding the absolute number of high-

quality anonymous fingerprints generated. From the

perspective of fast generation, StyleGAN is clearly

superior. ProgressiveGAN is preferable regarding the

better anonymity. Our future work will address the

training of GAN models based on multifarious

fingerprint images from many independent sources

and conditional generation of fingerprint patterns

such as predefined locations of minutia or substrate

characteristics.

ACKNOWLEDGEMENTS

This research has been funded in part by the Deutsche

Forschungsgemeinschaft (DFG) through the research

project GENSYNTH under the number 421860227.

We thank Philip Wiegratz, Alexander Heck and Mark

Trebeljahr for participating in this work at an early

stage and establishing the feasibility of GAN-based

generation of fingerprint images.

REFERENCES

Attia, M., Attia, M. H., Iskander, J., Saleh, K., Nahavandi,

D., Abobakr, A., Hossny, M., Nahavandi, S., 2019.

Fingerprint Synthesis via Latent Space Representation,

In Proc. IEEE Int. Conf. on Systems, Man and

Cybernetics, pp. 1855-1861.

Bontrager, P., Roy, A., Togelius, J., Memon, N.D., Ross,

A., 2018. DeepMasterPrints: Generating MasterPrints

for Dictionary Attacks via Latent Variable Evolution,

In Proc. 9th IEEE Int. Conf. on Biometrics: Theory,

Applications and Systems (BTAS), pp. 1-9.

Cao, K., Jain, A.K. 2018. Fingerprint Synthesis: Evaluating

Fingerprint Search at Scale. In Proc. Int. Conf. on

Biometrics (ICB), pp. 31–38.

Cappelli, R., 2009. SFinGe, In Li, S.Z., Jain, A. (eds)

Encyclopedia of Biometrics. Springer, Boston, MA

Cappelli, R., Lumini, A., Maio, D., Maltoni, D., 2007. IEEE

Trans. on Pattern Analysis and Machine Intelligence

29(9):1489-1503.

Elham, T., Olsen, M. A., Makarov, A., Busch, C., 2013.

Towards NFIQ II Lite: Self-Organizing Maps for

Fingerprint Image Quality Assessment, NIST

Interagency Report 7973, December 13, 2013.

Fahim M. A. I., Jung, H. Y., 2020. A Lightweight GAN

Network for Large Scale Fingerprint Generation, IEEE

Access, vol. 8, pp. 92918-92928.

Galbally, J., Cappelli, R., Lumini, A., Maltoni, D., Fierrez,

J., 2008. Fake Fingertip Generation from a Minutiae

Template, In Proc. 19th Int. Conf. on Pattern

Recognition (ICPR 2008), pp. 1-4.

Galbally, J., Haraksim, R., Ferrara, P., Beslay, L., Tabassi,

E., 2019. Fingerprint Quality: Mapping NFIQ1 Classes

and NFIQ2 Values, In Proc. Int. Conf. on Biometrics

(ICB 2019), pp. 1-8.

Garris, M., Mccabe, R., 2000. NIST Special Database 27

Fingerprint Minutiae From Latent and Matching

Tenprint Images, NIST Interagency Report 6534, June

1, 2000.

Hildebrandt, M., Dittmann, J., 2015. StirTrace V2.0:

Enhanced Benchmarking and Tuning of Printed

Fingerprint Detection, IEEE Trans. on Information

Forensics and Security 10 (4):833-848.

Karras, T., Aila, T., Laine, S., Lehtinen, J., 2017.

Progressive Growing of GANs for Improved Quality,

Stability, and Variation, CoRR abs/1710.10196.

Karras, T., Laine, S., Aila, T., 2018. A Style-Based

Generator Architecture for Generative Adversarial

Networks, CoRR, vol. abs/1812.04948.

Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J.,

Aila, T., 2019. Analyzing and Improving the Image

Quality of StyleGAN, CoRR, vol. abs/1912.04958.

Ko, K., 2007. User's Guide to NIST Biometric Image

Software (NBIS), NIST Interagency Report 7392,

January 21, 2007.

Minaee, S., Abdolrashidi, A., 2018. Finger-GAN:

Generating Realistic Fingerprint Images Using

Connectivity Imposed GAN. CoRR, vol. abs/

1812.10482.

Mistry, V., Engelsma, J.J., Jain, A.K., 2020. Fingerprint

Synthesis: Search with 100 Million Prints. CoRR, vol.

abs/1912.07195.

Orandi, S., Libert, J. M., Grantham, J. D., Ko, K., Wood, S.

S., Byers, F. R., Bandini, B., Harvey, S. G., Garris, M.

D., 2014. Compression Guidance for 1000 ppi Friction

Ridge Imagery, NIST Special Publication 500-289,

February 24, 2014

Ram, S., Bischof, H., Birchbauer, J., 2010. Modelling

fingerprint ridge orientation using Legendre

polynomials, Pattern Recognition 43(1):342-357.

Riazi, M.S., Chavoshian, S.M., Koushanfar, F., 2020.

SynFi: Automatic Synthetic Fingerprint Generation,

CoRR, vol. abs/ 2002.08900.

Wang, T., Liu, M., Zhu, J., Tao, A., Kautz, J., Catanzaro,

B., 2018. High-Resolution Image Synthesis and

Semantic Manipulation with Conditional GANs, In

Proc. IEEE Conference on Computer Vision and

Pattern Recognition (CVPR 2018), pp. 8798-8807.

Wyzykowski, A., Segundo, M., Lemes, R., 2020. Level

Three Synthetic Fingerprint Generation, CoRR, vol.

abs/ 2002.03809.

Zhang, H. Xu, T., Li, H., Zhang, S., Huang, X., Wang, X.,

Metaxas, D. N., 2016. StackGAN: Text to Photo-

realistic Image Synthesis with Stacked Generative

Adversarial Networks, CoRR, vol. abs/1612.03242.

Zinoun, F., 2018. Can a Fingerprint be Modelled by a

Differential Equation?, CoRR, vol. abs/1802.05671.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

352