High-speed Retrieval and Authenticity Judgment

using Unclonable Printed Code

Kazuaki Sugai, Kitahiro Kaneda and Keiichi Iwamura

Tokyo University of Science 6-3-1 Niijuku, Katusika-ku Tokyo, 125-8585, Japan

Keywords: Authenticity Judgment, Inkjet Print, Physically Unclonable Function, Feature Matching, Image Processing.

Abstract: The distribution of counterfeit products such as food packages, branded product tags, and drug labels, which

are easy to imitate, has become a serious economic and safety problem. To address this problem, we propose

a method to judge counterfeit products from commonly used inkjet-printed codes. To judge authenticity,

copies of printed matter of an inkjet printer are used as they are difficult to duplicate. In this study, we propose

a new authenticity judgment system that combines the locally likely arrangement hashing (LLAH) system,

which performs high-speed image retrieval, and Accelerated-KAZE (A-KAZE), which matches the features

of inkjet-printed matter with high accuracy to verify accuracy.

1 INTRODUCTION

In recent years, information and manufacturing

technologies have been advancing at an extremely

fast pace. While this has increased convenience in

daily life, determining the authenticity of information

and distinguishing between genuine and counterfeit

products have become increasingly difficult due to its

elaborate production. According to OECD/EIPO

(2019) survey the amount of counterfeit and pirated

goods traded worldwide trade is approaching $500

billion, which is equivalent to approximately 2.5% of

the world's total trade. Counterfeit goods range from

branded goods and airline boarding passes to

pharmaceuticals and food products. According to

WHO (2010) report a wide range of counterfeit

products, with the increased Internet penetration rate,

provide a dizzying array of both branded and generic

drugs. In more than 50% of cases, medicines

purchased online from illegal sites that conceal their

physical address were found to be counterfeit; many

of these products are labeled and packaged to look

genuine, making it difficult to distinguish between

counterfeit and genuine products at a glance. In

addition, counterfeit medicines have a tremendous

impact on society, not only in terms of economic

damage but also as a threat to the health and safety of

consumers, infringement of intellectual property

rights and trademark rights, and as a source of income

for illegal organizations.

In the field of security printing, holograms (Lim

et al., 2019) and security inks (Song et al., 2016) are

used, and artifact metrics using transmitted light

images (Matsumoto et al., 2000, Matsumoto et al.,

2004) have been used in Japan for some time.

Other methods have been proposed, such as the

use of unique random patterns created by

imperfections on the surface of objects, such as paper

plant fibers or industrial products (Yamakoshi et al.,

2007, Ito et al., 2004, Ito et al., 2018 and Fuji Xerox

Co., 2018), or the use of double-encoded codes, in

which the black cells of a barcode are encoded with

both normal and special black ink (Pacific Printing

Co., 2018).

In terms of productivity and cost, we think that the

inkjet printing method is best suited to the trend of

small-lot, multi-variety printing in the printing

industry, where the life cycle of products on the

market is short, consumer needs are diversified, and

high added value is sought. This research is focused

on the fact that the method is optimal in terms of

productivity and cost, and that it is possible to print

on various media such as film, cloth, plastic, and

wood, which is not possible with the

electrophotographic method.

This research aims to construct a system that

enables producers and consumers to judge product

authenticity by cross-verifying the printed codes on

labels and tags at the time of manufacture production

and delivery.

Sugai, K., Kaneda, K. and Iwamura, K.

High-speed Retrieval and Authenticity Judgment using Unclonable Printed Code.

DOI: 10.5220/0010542100570064

In Proceedings of the 18th International Conference on Signal Processing and Multimedia Applications (SIGMAP 2021), pages 57-64

ISBN: 978-989-758-525-8

Under a microscale, ink grains can be observed on

inkjet printouts. The shapes and positions of these

grains are slightly different for each print, and the

combination of these grains can be regarded as their

unique feature. Herein, we developed a system to

embed ink grains in a part of a printed code, which

can be enlarged and captured and verified that its

features (shape and position of the ink grains) are

registered in a database. The code can be

authenticated via highly discriminative color

matching.

Herein, we propose such a system by combining

Locally likely arrangement hashing (LLAH), a high-

speed image retrieval system that utilizes ink

variation in inkjet printing, and Accelerated-KAZE

(A-KAZE), which realizes the feature matching of

inkjet-printed materials.

2 PHYSICALLY UNCLONABLE

FUNCTION OF

INKJET-PRINTED CODES

The shape and positional relationship of the ink on the

paper can be a unique feature of materials printed by

an inkjet printer that can't be observed by laser

printer. The following factors can cause differences

in the shapes and positions of the ink.

1. Slight airflow on the paper surface

2. The direction in which the header moves when the

ink is fired

3. Air resistance of the ink in flight

4. Differences in the fiber quality of the adherend.

These properties can be used for authenticity

determination, as the uniqueness of these object

fingerprints (Physically unclonable function

properties), which each print possesses, makes it

difficult to reproduce maliciously. These properties

are independent of the printer type and apply to all

printers.

In this study, QR codes, typically used for airline

tickets and pharmaceutical labels, are used as an

embedding medium to obtain characteristics unique

to inkjet printing. Specifically, a single color

(grayscale information) with a density of 200 (Max

255), as shown in Figure 1, is embedded in the white

area in the upper left corner of the QR code to an

extent that it does not interfere with its use, and then

printed.

Figure 1: Embedded grayscale (density 200) (Max 255) and

QR code after embedding.

As an example, Figure 2 shows enlarged images

of three QR codes printed where the upper left white

area of each QR code was embedded with a single

color. But the same URL can be read from each QR

code.

Figure 2: QR codes and an enlarged photo of the upper left

corner.

The figure shows that, for the printed QR codes

with embedded grayscale information, the shape and

position of the printed ink differ greatly, and the same

ink arrangement can be difficult to reproduce exactly

on a microscopic level. Therefore, the characteristics

of printed material can be treated as features for

authenticity determination. And this study has shown

that these features are not lost after three years of

storage.

3 METHOD OF HIGH-SPEED

RETRIEVAL AND

AUTHENTICITY JUDGMENT

ALGORITHM

3.1 Overview of the System

The outline of the proposed system is shown in Figure

3, which can be classified primarily into the following

processes.

1. Capturing the QR code and saving the image

2. Image preprocessing for LLAH

3. Fast retrieval by LLAH

4. Authenticity judgment by feature matching using

A-KAZE

SIGMAP 2021 - 18th International Conference on Signal Processing and Multimedia Applications

Figure 3: Outline of the proposed new system for determining authenticity.

Figure 4: The capturing environment used and the actual

capturing scene. Left: The phone and microlens used.

Right: The image capture set up.

3.2 Image Preprocessing

We applied image preprocessing to extract feature

points for the LLAH system using the captured

images. The following procedure was used to

preprocess the image and produce a binary image in

which the ink hues were extracted from the grayscale

part by dividing it into three colors: cyan, magenta,

and yellow. Each color was extracted by setting a

range in the Hue, Saturation, Value (HSV) color

system. Which is calculated by using OpenCV3.30

library with threshold value.

1. Hue

Cyan: 90 to 150,

Magenda: 150 to180,

Yellow: 20 to 40

2. Saturation All: 90 to 255

3. Value All: 120 to 255

Figure 5 shows the cropped image immediately after

capturing and the ink colors extracted. (Colors are

added for clarity).

Figure 5: Comparison before and after preprocessing.

The process for image preprocessing is as follows.

1. Cropping of the input image to 100 × 100 pixels

along the right and bottom edges.

2. Shrinkage processing.

3. Detect the outline of the QR code and the

boundary line between black and white.

4. Determine the angle of the above boundary line

and correct the angle of the captured QR code.

5. The origin of the boundary line is calculated using

a new algorithm that adds assistance to the

approximation.

6. Crop the image further to the size of 700 × 500

pixels.

7. Adjust contrast.

8. Extract the three ink hues.

9. Grayscale conversion.

10. Adaptive Binarization.

11. Smoothing by Gaussian filter.

12. Re-binarization.

13. Invert bits to create an image for input into LLAH.

14. Shrink the image again to remove the noise.

3.3 Fast Retrieval by LLAH

The feature matching system A-KAZE was

introduced into the proposed system to judge the

authenticity of printed matter. Even though the

system is very accurate in judging the authenticity of

inkjet-printed matter, it takes approximately 0.4 to 1.0

s for a one-to-one match. To match 100,000 images

with Only A-KAZE, the system would take

approximately 27 h, which is far from our goal within

10s ,so it is impractical for a real system. Therefore,

in the proposed system, we used a technique called

locally likely arrangement hashing (LLAH), which

was proposed by Nakai et al., 2006, Iwamura et al.,

2007, and Takeda et al., 2011, that can retrieve

matching images as quickly as 0.02 to 0.04 s

regardless of the number of images registered in the

database.

Previously, the LLAH technique was used for

document image retrieval. Document image retrieval

is a technique where images similar to the input image

are retrieved. For example, by taking a picture of a

LLAH

rieval

A-KAZE

ching

Real

Fake

Feature point

extraction

Calculation of

features

High-speed Retrieval and Authenticity Judgment using Unclonable Printed Code

paper article with a digital camera and inputting it into

the retrieval system, the user can retrieve pre-

registered documents and their references. Web

services for this technique are also available.

LLAH does not use the image information as is

but a unique index value obtained from the image for

the registration and matching process, thus reducing

the data size and realizing high-speed and high-

performance retrieval.

3.3.1 LLAH Extraction of Feature Points

and Calculation of Feature Values

Since LLAH judges the match of images based on the

arrangement of feature points, the same feature points

must be extracted even in the case of projective

distortion, noise, or low resolution. Therefore, the

center of gravity is used as a feature point and

obtained from the ink portion of the binary image

generated during preprocessing.

The positions of the obtained feature points are

indicated by coordinate information. These positions

will change with the angle during capturing changes.

To obtain the same feature values even in such a

situation, the feature values are not calculated from a

single feature point but calculated based on the

positional relationship between feature points. In

addition, since the images captured by the camera are

subject to projection transformation, we used affine

invariants as the feature values, which was proposed

by Nakai et al., 2006, and which are robust to changes

in the position of the feature points as geometric

invariants robust to such changes. Figure 6 is shown

to explain affine invariants.

Figure 6: Calculation of affine invariants.

When there are four points A, B, C, and D on the

same plane as shown in the above figure, the affine

invariant for point A is calculated by the following

equation:

𝑃

𝐴

,𝐶,𝐷

𝑃

𝐴

,𝐵,𝐶

(1)

where P (A, B, C) is the area of the triangle formed

by the three points A, B, and C. This feature is

calculated and discrete-valued from the nearest point

to a point. When obtaining m points in the vicinity of

a certain point, it is simplest to set m = 4; however, if

this is the only method used, the same feature value

may be calculated for features in different images and

result in an incorrect match. Therefore, to further

improve the discriminability, we can set m ≥ 4, find

all combinations of four points out of m points, and

use the sequence of discretized values of affine

invariants computed from the combinations as feature

values. In this study, we set m = 6. Each sequence was

uniquely determined by moving in a clockwise

direction.

3.3.2 Registration with the Database

When registration the extracted features in the

database, we obtained the index, H

index

, and

quotient, Q, from the extracted features so that they

can be uniquely retrieved using the following

equation:





𝑟



𝑖



𝑚

𝐶

1

𝑖0

𝑑

𝑖



𝑄𝐻

𝑠𝑖𝑧𝑒

𝐻

𝑖𝑛𝑑𝑒𝑥

(2)

where r_(i) is the discrete value of the i-th affine

invariant, d^i is the number of discrete value levels,

and H

size

is the size of the hash table. In this study, we

set d = 10 and H

size

= 1086. In other words, in two

matching images, the same index, H

index,

and quotient

Q can be obtained from the same feature vector.

Therefore, during retrieval, for features with the same

index

(when collisions occur at the same index),

quotient Q is used for comparison instead of the

feature values. Therefore, as shown in Figure 7, the

image ID, feature point ID, and quotient Q are

registered in the database. The image ID and feature

point ID are the image and feature point identification

numbers, respectively. When a collision occurs in the

index, a list is added.

Figure 7: Configuration of the database (Takeda et al.,

2006).

SIGMAP 2021 - 18th International Conference on Signal Processing and Multimedia Applications

Figure 8: Comparison of registered data with collation data (right) and fake data (left).

3.3.3 Retrieval from Database

Matching is performed by voting on registered

images using a voting table. First, as in the

registration process, the index of the database is

calculated for each feature point, and the obtained

index is used to refer to the list shown in Figure 7. For

each item in the list, we checked if the quotient

matches, and if it does, we increment the image ID

item in the voting table. Finally, the results were

sorted by the number of increments, and the images

were outputted as the retrieved results.

3.3.4 Need for Additional Feature Matching

By setting a threshold value at the "Retrieval from

database" stage in the LLAH system introduced in

Section 3.3, the system can determine the authenticity

of collation and fake data. However, since the LLAH

system uses binary black and white images for

matching, a high threshold value is required to reject

all fake data when the dataset is large. Therefore, we

proposed combining color matching with LLAH to

make a more powerful system for judging

authenticity, instead of using LLAH alone.

3.4 Judgment of Authenticity by

A-KAZE

To solve the above problem, we introduced A-KAZE

matching, which can perform highly discriminative

verification using color images, while LLAH, which

can independently determine authenticity, is used for

verification from the database without setting a

threshold.

A-KAZE was introduced (Tareen et al., 2018,

Chien et al., 2016) as a feature matching method that

has high robustness, similar to SIFT and SURF, and

is good at extracting feature points on a plane;

therefore, is resistant to rotation and illumination

changes. Figure 8 shows the results of the A-KAZE

matching. Figure 8 (right) shows the comparison of

registered data collation data using A-KAZE, and

Figure 8 (left) shows the comparison of registered and

fake data. As can be seen, more than 200 features

were matched to the collation data, whereas none

were matched to the fake data. In addition, the

maximum number of matches for the fake data was

only four points, and the genuine data matches were

over 50 times higher.

However, although the feature matching A-KAZE

is very accurate for inkjet-printed materials, as

mentioned earlier, it takes approximately 0.4 to 1.0 s

for one-to-one matching, and when applied to an

assumed scale of 100,000 sheets, it takes

approximately 27 h per sheet. This is not practical for

real systems. Therefore, in this system, we will

continue to use the LLAH system, which is capable

of high-speed retrieval. So, we can accurately judge

authenticity within 10s, which is our target time and

does not burden the user. In the proposed system in

this research, the feature matching of A-KAZE is

introduced as follows: For each image divided into

three colors (cyan, magenta, and yellow), the top

three candidates for each color are retrieved using

LLAH, and a total of nine candidates (three colors ×

three candidates) are selected. We then perform A-

KAZE matching on each of the three candidates per

color to obtain a highly accurate judgment of

authenticity.

4 VERIFICATION

4.1 Verification Outline

In this study, we printed 2,000 QR codes for

verification. The images were taken in an

environment using a smartphone as described in "3.2

QR Codes Capturing." The captured images were

categorized as follows:

1. Registration data: data taken for registration of

1000 QR codes.

2. Collation data: The same 1000 QR codes were re-

captured and used to match with the registration

images.

3. Fake data: data taken from the second 1000 QR

codes without registration.

A total of 3000 images were collected in this manner.

To show that the method does not depend on the

environment in which the images are taken, the ima-

High-speed Retrieval and Authenticity Judgment using Unclonable Printed Code

ges were taken on different days.

To verify the accuracy of the judgment of

authenticity, we changed the size of the datasets

registered in the database. Specifically, verifications

were conducted in four dataset sizes: 20, 100, 400,

and 1000 registered, collation, and fake data.

Furthermore, the effectiveness of the proposed

method combined with A-KAZE matching was

verified by comparing the results achieved using only

the LLAH system with those using LLAH with A-

KAZE and to compare with other feature detectors, a

comparison with ORB is also shown to demonstrate

the superiority of A-KAZE.

4.2 Equipment

The following list shows the instruments used for the

verification

1. Printer:

IP8730 (print resolution 9600dpi) by Canon

Japan

2. Smartphone:

iPhone XR (camera resolution 12 million pixels)

by Apple America

3. Microlens:

i micron pro (maximum magnification 800x) by

Qing Ting E&T LLC Japan

4. Computer:

Magnate IM (Intel Core i5-9400 @ 2.90GHz,

16GB RAM) by THIRDWAVE Japan

5 RESULTS

5.1 Verification Results

Table 1 shows the results of the verification of

authenticity for each dataset size. The upper part

shows the results of the verification of the

authenticity using only the LLAH system, LLAH

with ORB and the lower part shows the results of the

proposed authenticity determination system

combined with A-KAZE.

Table 1 shows that the correctness rate for the

collation data was very high, exceeding 98%, for all

dataset sizes. In addition, the proposed system is

sufficiently accurate to be used in the real world.

Furthermore, a comparison of the accuracy of the

proposed system with that of the LLAH system alone

showed that the accuracy of the proposed system was

over 95% for small datasets of up to 400 images.

However, when A-KAZE was included, the accuracy

was further improved for the small dataset of up to

400 images, and it was confirmed that the accuracy

did not decrease and remained high even when the

number of images were increased to 1,000, which was

a point of concern.

The high accuracy can be attributed to the LLAH

being used only for the retrieval function without a

threshold value, which is a key feature of the new

method. This allowed data to be considered positive,

where previously it would have been excluded

because the threshold value was not reached when

judged by LLAH alone.

Furthermore, the accuracy of the A-KAZE

matching system was confirmed to be 100% for

candidates of the collation and fake data, given as

arguments. On the other hand, compare with the

accuracy of LLAH with ORB, one of the feature

detectors for corner, can be seen that there is almost

no improvement in accuracy, and instead the

accuracy decreases.

In summary, the results confirm that A-KAZE is

a suitable feature matching system for the authenticity

determination of the inkjet-printed matter.

As for the average processing time, the following

breakdown was obtained for 1000 registered images.

LLAH (retrieval function): 0.25 s.

A-KAZE (authenticity judgment function): 8.42 s

Total: 8.67 s

Therefore, this system enables fast retrieval of images

by LLAH and judge for high accuracy authenticity by

A-KAZE within 10seconds, which is the target time

and fast enough for practical use. In the case of ORB,

the speed is 0.89s, which is very fast and slow

compared to that, but this is the target time and is fast

enough for practical use.

In addition, the storage usage of the system with

1000 registered images was as follows:

LLAH (retrieval function): 87 MB

A-KAZE (authenticity judgment function): 492 MB

Total: 579 MB

This breakdown shows that the database for A-KAZE

matching is approximately 6.6 times that used for

LLAH alone because it uses color images as they are.

Therefore, in real-world applications, the color

images for A-KAZE should be stored on a server in

the cloud rather than the user terminal, and responses

should be sent in response to requests.

5.2 Discussion

From the results of the verification described in

Section 5.1, we verified that if the correct candidate

is passed to A-KAZE as an argument, the system can

determine the authenticity with 100% accuracy.

SIGMAP 2021 - 18th International Conference on Signal Processing and Multimedia Applications

Table 1: Judgment results for each dataset size of the authenticity judgment system (Numbers in red show where the accuracy

of the proposed method was better than LLAH only and LLAH with ORB).

Therefore, observed that the accuracy of LLAH is

directly related to the correctness of the judgment.

Table 2 summarizes the results of the first, second,

and third candidate retrievals for cyan, magenta, and

yellow for a total of three sets of 1000 images of

registration data.

Table 2: Retrieval accuracy by color (1000 registered data).

The table shows that the result for the three sets

combined is 98.6%, which is very high, but the result

for the authenticity judgment of each color is between

85% and 90%, which means that there is room for

improvement. This is because when extracting the

inks in the image preprocessing stage, the specular

reflection of the flash during capturing prevents

accurate extraction of some colors, resulting in

misalignment of the center of gravity and different

feature values, leading to matches with more images.

Figure 9 shows three examples, one of which was

retrieved correctly, whereas the other two failed.

Figure 9: Comparison of successful and unsuccessful

images for determining authenticity.

The correctly retrieved image (left) is clear, but

the failed images have problems, such as red due to

the flash of the smartphone (center) and blue (right).

For the practical use of the system, numerical analysis

and verification of the images must be analyzed and

verified before image preprocessing. This will

prevent registration and judgment of authenticity on

images where it is difficult to extract ink colors,

obtain the center of gravity, and calculate feature

values.

High-speed Retrieval and Authenticity Judgment using Unclonable Printed Code

6 CONCLUSIONS

In this study, we proposed a high-speed and high-

accuracy authentication system by combining LLAH,

which can perform high-speed image retrieval, and

A-KAZE, which can perform high-accuracy feature

matching for inkjet-printed matter.

Our evaluation verification using a large dataset

of 1,000 images showed a correct judgment rate for

collation data exceeding 98%, and a positive

judgment rate of 100% for fake data within 10s,

which is exceptional even considering its use in the

real world.

As mentioned in the Discussion section, there is

still room for further improvement in accuracy of the

retrieval system. Further, the image preprocessing

needs to be enhanced to improve the accuracy of the

entire system.

In addition, we consider the introduction of a

verification step before the preprocessing stage to

detect images that are unsuitable for registration and

authenticity judgments (Figure 9) by getting the

average hue of the image. and to prompt the user to

re-capture the image.

REFERENCES

OECD/EIPO. (2019). Trends in trade in counterfeit and

pirated goods, Illicit Trade. OECD Publishing Paris

[PDF file], Retrieved from https://www.oecd-

ilibrary.org/trade/trends-in-trade-in-counterfeit-and-

pirated-goods_g2g9f533-en. (accessed 2021-2-10)

The World Health Organization (WHO). (2010). Growing

threat from counterfeit medicines. Bulletin of the World

Health Organization, Vol. 88, (4), April 2010, pp.247-

248.

Lim, K.T.P., Liu, H., Liu Y., and Yang J.K.W. (2019).

Holographic color prints for enhanced optical security

by combined phase and amplitude control. Nature

communications, Vol. 10, (25).

Song, Z., Lin, T., Lin, L., Lin, S., Fu, F., Wang, X., and Guo,

L. (2016). Invisible Security Ink Based on Water-

Soluble Graphitic Carbon Nitride Quantum Dots.

Angewandte chemie, Vol. 128, (8), pp.2823-2827.

Matsumoto, H., Matsumoto, T. (2000). Artifact-metric

Systems. Technical Report of IEICE, ISEC 2000-59,

pp.7-14.

Matsumoto, T., Iwashita, N. (2004). Financial Operations

and Artifact Metrics. Financial Studies, Vol. 23, (2), pp.

153-168. (in Japanese)

Yamakoshi, M., Tanaka, J., Furuya, M., Hirabayashi, M.,

Matsumoto, T. (2012). Individuality Evaluation for

Paper Based Artifact-Metrics Using Infrared

Transmitted Light Image. JAPAN TAPPI JOURNAL,

Vol. 66, (4), pp.419-425.

Ito, K., Shimiz.u, T., Sugino, T. (2004). Paper Identification

and Matching Device and Method. Japanese Patent

Publication 2004-102562.

Ito, K., Hayashi, E. (2018). Technology for the Recognition

of Unique Objects: Yoctrace®. Fuji Xerox Technical

Report 2018, (27), pp.79-85. (in Japanese)

Fuji Xerox Co.,Ltd (2018). Technology for the Recognition

of Unique Objects: Yoctrace®.

https://www.fujixerox.com/eng/company/technology/c

ommunication/annotation/yoctrace.html (accessed

2021-2-19)

Pacific Printing Co.,Ltd. (2018). Fake Code Buster: A

Proposed Solution for Determining Authenticity.

http://www.busi-tem.jp/code/code.html/. (accessed

2021-2-19) (in Japanese)

Nakai, T., Kise, K., and Iwamura, M. (2006). Use of Affine

Invariants in Locally Likely Arrangement Hashing for

Camera-Based Document Image Retrieval. Document

Analysis Systems, Vol. 7, pp541-552.

Takeda, K., Kise, and Iwamura, M. (2007). Improvement

of Retrieval Speed and Required Amount of Memory

for Geometric Hashing by Combining Local Invariants

In proc. BMVC’07.

Iwamura, M., Nakai, T., Kise, K. (2011). Real-Time

Document Image Retrieval for a 10 Million Pages

Database with a Memory Efficient and Stability

Improved LLAH. International Conference on

Document Analysis and Recognition 2011.

Tareen, S.A.K., Saleem, Z. (2018). A Comparative

Analysis of SIFT, SURF, KAZE, ORB, and BRISK.

International Conference on Computing, Mathematics

and Engineering Technologies 2018

Chien, H.J., Chuang, C.C., Chen, C.Y., and Klette, R.,

(2016). When to Use What Feature SIFT, SURF, ORB,

or A-KAZE Features for Monocular Visual Odometry,

International Conference on Image and Vision

Computing New Zealand 2016

SIGMAP 2021 - 18th International Conference on Signal Processing and Multimedia Applications