product piracy prevention: product counterfeit detection...

Product Piracy Prevention: Product Counterfeit Detection Without SecurityLabels

Christian Horn1, Matthias Blankenburg2, Maximilian Fechteler1 and Jorg Kruger21Department of Industrial Automation Technology at Berlin Institute of Technology

2Fraunhofer Institute for Production Systems and Design [email protected], [email protected], [email protected]

[email protected]

ABSTRACT

Counterfeiting is a challenge to companies, cus-tomers and markets all over the world. Besidesthe economic damage which affects in particularthe companies and countries that use advanced pro-duction and manufacturing processes based on in-tensive research and development to produce highquality goods, safety standards are omitted. Thesestandards protect usually the customer from goodswhich are dangerous or harzardous to health.Product piracy prevention is often followed bythe application of RFID tags to supervise supplychains. The lack of robust counterfeit detectionmethods created a market for artificial security la-bels which are used to secure the product itself.The specific conditions of production, manufac-turing technologies and materials generate specificfeatures, which identify every product uniquely.The innovation of this text is the detection of thesefeatures in an automated fashion through the com-bination of digital sensing and machine learning,rendering the application of artificial security la-bels obsolete.

KEYWORDS

Automated Counterfeit Detection, Product Finger-printing, Pattern Recognition, Sensor Fusion, Clas-sification

Figure 1: Cases of customs enforcements of intellec-tual property rights at the european border, from [1]

1 INTRODUCTION

Figure 1, taken from the annual ”Report onEU customs enforcement of intellectual prop-erty rights” of the European Union in 2012[1], shows a continious upward trend in thenumber of shipments suspected of violatingintellectual property rights for the last years.In 2011 more than 90 thousand cases of de-tained articles were reported. The value totheir equivalent genuine products is estimatedto be over 1.2 billion euro and this covers onlyEurope.To get an idea of the worldwide amount ofeconomic damage for the last years the re-port ”The Economic Impact of counterfeit-

International Journal of Cyber-Security and Digital Forensics (IJCSDF) 2(1): 88-102The Society of Digital Information and Wireless Communications, 2013 (ISSN: 2305-0012)

88

ing and piracy” [2] of 2008 estimates a totalloss of 250 billion dollars in the year 2007.This report covers the analysis of internationaltrade in counterfeit and pirated products, butthese estimates do not include domesticallyproduced and consumed counterfeit and pi-rated digital products being distributed via theInternet. If these were also considered, themagnitude of counterfeiting and piracy world-wide could be several hundred billion dol-lars more in 2007. Furthermore, if we com-pare these numbers to the amount of cases re-ported in Figure 1, they probably doubled in2011. The effect of counterfeiting and piracyis an intermission of innovation and thus im-pairment of economic growth. The economicdamage affects in particular countries that useadvanced production and manufacturing pro-cesses based on intensive research and devel-opment to produce high quality goods.Another very important argument to enablethe differentiation between brand products andtheir counterfeits is safety. It is stated in theOECD report that the products counterfeitersand pirates produce and distribute are often ofminor quality and can even be dangerous andhealth hazards. Common standards that ensurethe safety of products can be ignored by prod-uct pirates and the used materials can be dan-gerous.With the magnitude of counterfeiting andpiracy in mind, these reports emphasize theneed for more effective enforcement to com-bat the counterfeiting and piracy on the partof governments and businesses alike. A keycomponent for this enforcement is the devel-opment of new methods for automated coun-terfeit detection.The review of copyright infringement of reg-istered trademarks and products is not easy toimplement. Due to the high number of pend-ing trademarks and constantly added new ap-plications it is very difficult for the executivebodies, such as customs, to register violations

Figure 2: Current scenario for counterfeit detectionthrough customs officials

Figure 3: Desired scenario for counterfeit detectionthrough customs officials

of trademark rights immediately and in a com-prehensive manner. The awareness to all reg-istered brands and products is for the execu-tive organs not possible and therefore neces-sarily, trademark infringement remains unno-ticed. The current scenario for products enter-ing a market in a foreign country is diplayed inFigure 2. Here it is shown how customs offi-cials usually handle the inspection of productsat the border. First the goods arrive at a spe-cific check point, usually via sea- or airfreight.If the customs officer notices some anomaly inthe paperwork, he will check the cargo con-tainers. As discussed earlier the officer isoften not an expert for the shipped product,so he could not detect a counterfeit. Insteadthe company producing the genuine product iscontacted to send their own expert, which canverify the product. This is a time-consumingand expensive process, therfore most contain-ers in question often remain unnoticed.To overcome these limitations in the checkuproutine an automated expert-system is neces-


89

sary that can support the customs officials, asshown in Figure 3. Given that the officer couldverify the shipped cargo by himself while thecompany issues the authentication system fortheir products. This idea was adopted morerecently through an application of artificial se-curity features to products. The issues of suchsecurity labels are in part the high cost, andadditionally the integration into the product.On the other hand high-quality branded prod-ucts, as the target of counterfeiting, have usu-ally, due to the production processes and ma-terials used, and in view of its processingmachinery and equipment, a grade of highquality. The specific conditions of produc-tion, manufacturing technologies and materi-als generate specific features, which identifythe product uniquely. These features may bedetected multimodal by man, including tac-tile (plasticity, elasticity, thermal conductivity,surface structure), visual (shape, colour, sur-face texture, transparency), olfactory (smell)or acoustic (sound) perceptions. In general,only the person familiar with the manufac-ture of the product can combine these inher-ent characteristics in their entirety so that itcan differentiate the genuine product from aclear counterfeit. The innovation of this textis the detection of these features in an auto-mated fashion through the combination of dig-ital sensing and machine learning, renderingthe application of artificial security labels ob-solete. As shown in Figure 4, more than 60%of counterfeit products are shoes, bags andclothing. Therfore two properties of a prod-uct have been identified as the most promisingones suitable for identification: the olfactoryand the optical features.

2 STATE-OF-THE-ART-TECHNOLOGY

Common automated counterfeit detectionmethods require nowadays additional security

Figure 4: Categories of counterfeit products, from [1]

features at the product itself. Several methodshave been developed, but main advantages anddisadvantages remain similar.Additional security features require furthersteps in production to add these features to theproduct. This raises expenses, manufacturingtime and development efforts, which is clearlya disadvantage. On the other hand the securityis enhanced and an original brand is easy todetect in an automated fashion, since there is aspecific feature to look for. But this could alsobe a main disadvantage, if the security featureitself is easy to reproduce and could be addedto any forged product. Another challenge isto link the securtiy label to the brand productin a way it cannot be removed or stolen. Thisway product pirates could label their counter-feits easily as an original with an original se-curity label.Figure 5 shows examples of different labelswhich are commonly used on products for dif-ferent purposes. One purpose is the use as alogical security feature where the security la-bel contains unique information and cannot becopied. Counterfeit detection without artificialsecurity tags is a solution to these problems,if the counterfeit is distinguishable from theoriginal brand.

2.1 Security Labels

The report [3] of the German Engineering Fed-eration shows the latest innovations against


90

(a) QR-Code (b) RFID

(c) FraunhoferSecuriFlex

Figure 5: Different types of security labels

product piracy. It gives a comprehensiveoverview of the latest efforts in product pro-tection. A reasonably well studied approachis the extensive supervision of supply-chains.Here the application of RFID tags plays a sig-nificant role, as the latest form of artificialsecurity tags, which can easily be integratedwith existing logistic chains. The applica-tion of Data Matrix Codes (DMC) is discussedas well as a cost-effective alternative. Muchwork has been done to link these tags insepa-rably with the corresponding product to hinderproduct pirates from transferring these tags totheir counterfeits. But in general it is observedthat this protection method holds only withtremendous logistic implications, since todaysproducts cover various stations during the dis-tribution process. Up to now there has been nocommon standard available and the customsauthorities’ integration is still open. Evenwhen the cost of these artificial tags could bereduced by advances in the production pro-cess, as e. g. the introduced direct printing ofRFID antennas onto packaging, additional ex-penses with no direct use for the customer willarise. Security Tags like holograms found at-tached to various consumer goods give nearlyno protection against counterfeiting since ma-chine readability is poor and knowledge of the

correct appearance is scarce.

2.2 Product-Inherent Features

The Inherent ID Project adopts a novel ap-proach to protecting high-value products fromcounterfeiting. The approach is based on thestationary and mobile capture of key prod-uct features indissolubly linked with the prod-uct which enable its production process to betraced. This not only renders obsolete the ap-plication of security tags but also gives en-hanced protection against counterfeiting as theinherent characteristics that the high-qualityproduction process impregnate in the genuineproduct are combined with one another toserve as proof of product identity. They formthe basis on which electronic certificates of au-thenticity can be issued without the need forcomplicated explicit security markings. Meth-ods for the capture and control of identity char-acteristics are being elaborated in the InherentID project for system integration using intel-ligent cameras and an electronic nose. Theidentity characteristics captured by this rangeof sensors serve both for the product identi-fication and product authentication. At thesame time this also offers opportunities for im-proving documentation of product flows in thesupply chain. Full documentation serves as acomplement to the inherent characteristics ofthe authentic product and offers valuable in-formation of verification of the genuine arti-cle, thus serving to safeguard against counter-feits. The Project aimes to answer the ques-tion: Which inherent features allow separationof genuine products from counterfeits in an au-tomated fashion? The motivation of this ques-tion is the assumption that genuine productsmust differ in its properties from its counter-feit, since the product pirate tries to maximizeits profit by using material of inferior qualityand misusing a trademark of a genuine man-ufacturer to feint the customer. One result of


91

Figure 6: Concept of the Project Inherent ID

the project is that only a combination of fea-tures can detect counterfeits at a decent ratefor different products.

3 MATERIALS AND METHODS

Optical 2D and 3D characteristics as well asolfactory characteristics are combined withone another to serve as proof of product iden-tity, as shown in Figure 6. They form the basison which electronic certificates of authentic-ity can be issued without the need for com-plicated explicit security markings. The iden-tity characteristics captured by this range ofsensors serve both for product identificationand product authentication. At the same timethis also offers opportunities for improvingdocumentation of product flows in the supplychain. Within the scope of Inherent ID is thesuccessful establishment of a laboratory pro-viding multi-modal measurement equipmentcomprising multigas sensor array for olfactoryanalysis, high resolution camera for textureanalysis and stereo vision, as well as rangecameras for 3D feature extraction. Further re-search is conducted with the aim for increas-ing robustness of the sole test methods espe-cially under ambiguous environments, integra-tion into portable devices, implementing sen-

sor data fusion for increased detection ratio,effortless integration into supply chains anddeveloping efficient data models for storageof various features depending on the regardedproduct.

3.1 Texture Features

The ability to characterise visual textures andextract the features inherent to them is consid-ered to be a powerful tool and has many rel-evant applications. A textural signature capa-ble of capturing these features, and in partic-ular capable of coping with various changesin the environment would be highly suited todescribing and recognising image textures [6].As humans, we are able to recognise textureintuitively. However, in the application ofComputer Vision it is incredibly difficult to de-fine how one texture differs from another. Inorder to understand, and manipulate texturalimage data, it is important to define what tex-ture is. Image texture is defined as a functionof the spatial variation of pixel intensities [5].Furthermore, the mathematical description ofimage texture should incorporate, identify anddefine the textural features that intuitively al-low humans to differentiate between differ-ent textures. Numerous methods have beendesigned, which in the past have commonlyutilised statistical models, however most ofthem are sensitive to changes in viewpoint andillumination conditions [6]. For the purposesof mobile counterfeit detection, it is clear thatthis would be an important characteristic forthe signature to have, as these conditions cannot be entirely controlled. Recently a descrip-tion method based on fractal geometry knownas the multifractal spectrum has grown in pop-ularity and is now considered to be a use-ful tool in characterising image texture. Oneof the most significant advantages is that themultifractal spectrum is invariant to the bi-Lipschitz transform, which is a very general


92

UserInput

Gaussian Smoothing

Gradient Filtering Laplacian Filtering

Calculate LocalFractal Dimension

Define Setsfrom LFD Map

Calculate GlobalFractal Dimension for Sets

TextureSignature

Figure 7: Workflow for generating a textural signature

transform that includes perspective and texturesurface deformations [6].Another advantage of Multrifractal Spectra isthat it has low dimension and is very efficientto compute [6] in comparison to other methodswhich achieve invariancy to viewpoint and il-lumination changes such as those detailed in[7], [8]. One of the key advantages of mul-tifractal spectra, which is utilised here is thatthey can be defined by many different cat-egorisations or measures, which means thatmultiple spectra can be produced for the sameimage.This is achieved through the use of filtering,whereby certain filters are applied to enhance

1 1.5 2 2.5 3 3.50

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

Local fractal dimension α

Glo

bal f

ract

al d

imen

sion

f(α)

GaussianGradientLaplacian

1 1.5 2 2.5 3 3.50

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

Local fractal dimension α

Glo

bal f

ract

al d

imen

sion

f(α)

GaussianGradientLaplacian

Figure 8: Multi-Fractal-Spectra of texture of a textileproduct (top) and its counterfeit (bottom)

certain aspects of the texture, to create a newmeasure. Certain measures are more or less in-variant to certain transforms, and the combina-tion of a number of spectra achieves a greaterrobustness to these. The worklfow is depictedin Figure 7 and an example is given in Fig-ure 8.

3.2 Shape Features

Since manual detection is often done visualby customs officials, visual features are alsoimportant for any automatic detection mecha-nism. Besides detecting features through twodimensional image processing, three dimen-sional data capture is necessary for counterfeitdetection, because it provides important addi-tional information.To capture a real-world object in three di-mensions a 3D scanner, or range camera, canbe used. The basic principles of 3D scan-ners available on the market are triangulation,time-of-flight or interferometric approaches,whereas each principle has its advantages ordisadvantages. For a profound insight into that


93

Figure 9: Scanned shoes from our database

topic refer to [9]. We use a mobile structured-light 3D scanner for our application, but ingeneral any three dimensional data acquisitionmethod can be used to capture a real-world ob-ject. But while using different kinds of scan-ning techniques the results may vary.One distinguishable feature of brand productsis the shape itself. Shape matching is a wellstudied topic and several publications can befound over the last 15 years. Despite manydifferent approaches available, most practicalapplications still use the 1992 introduced Iter-ative Closest Point Algorithm (ICP) [10] or itsoptimized variants to match objects. This isdue to the fact that most newer approaches areneither easy to implement nor able to run ata reasonable speed for the use in commercialsoftware.One major challenge for three dimensional ob-ject capture is the huge amount of data that hasto be processed. The 3D scanner we use has anaccuracy of 20 to 50 µm and generates around300, 000 vertices per object. Assuming a pointper point matching algorithm with O(nc) andc > 1 growth rate and a calculation time of1ms per point match, it would take nearly 3years to calculate a match of two objects.Feature-based approaches have become verypopular since some years in image analysis(2D) due to robustness and less computationaleffort compared to other approaches. In shapematching (3D) feature-based approaches havebeen introduced more recently and are gainingpopularity in shape retrieval applications forthe same reasons. The major difference amongthese is whether the approach uses global orlocal features. A global feature describes the

(a) 1/f-Noise (b) White Noise

Figure 10: Different kinds of noise applied to a mesh

whole object, while local features only de-scribe parts or details of an object. In [11] anoverview of shape matching principles and al-gorithms can be found.Many shape matching approaches use digitalhuman made data like the Princeton-Shape-Benchmark [12] or the SHREC datasets[13] to evaluate their algorithms. Scanneddata from real world objects is different toarificially-made data in a sense that holes1 andvariations between two scans of the same ob-ject can appear. The SHREC datasets haveindeed several categories with different 3Dmodels to mirror these real-world challenges,but the categories are examined separately andmodels are artificially-made too. For thatreason we created our own database using3D scanners and our students shoes. Fig-ure 9 shows some examples of our scans. Wescanned some shoes with different scanners toget a more complex testing database. Further-more different types of noise were applied tothe scanned models as shown in Figure 10.Approaches using global features are not suit-

1Holes are areas on the scanned object where theused scanning technique has troubles to capture data.


94

Figure 11: Laplacian Smoothing

able for counterfeit detection, where minordetails of an object can be highly important.Therefore only approaches detecting local fea-tures were taken into consideration. Our auto-matic local-feature-based matching algorithmconsists of two major parts: a feature detectorand a feature descriptor. The classification isdone later after the texture and odour featuresare combined with the shape features troughfeature fusion.

Feature Detector

The feature detector finds points of inter-est on a given mesh which are usually ex-trema in a specific mathematical notation. Intwo-dimensional approaches well known tech-niques like corner detection are used. In threedimensions new approaches based on two-dimensional image processing algorithms thatuse feature-based approaches have been de-veloped. Examples are the Harris-3D-featuredetector [14], several portations of the SIFT-algorithm to three dimensions [15, 16] orthe 3D equivalent of SURF [17]. Otherapproaches use for example Heat-Kernel-Signatures [19] or maximally stable extremalregions (MSER) [18] to detect features.For countereit detection we use a Scale Spaceapproach to detect keypoints [9]. The ScaleSpace is usually constructed by repeatedly ap-plying a filter to a given mesh.

L(x, y, z, σ) = F (x, y, z, α) ∗M(x, y, z)

whereas M ist the mesh and F ist the filter-kernel. The difference of the resulting meshes

Figure 12: Key Points for two Scans of the same Shoe

is then examined for extremas. As filter-kernel a finite difference approximation of theLaplace operator

G(x, y, z, α) =1

n

n∑i=1

αiPi

was used, where αi is a weighting factor andPi are the neigbors of the regarded point. Theadvantage of this smoothing approach is thateach point keeps its relative position, keepingthe shape itself of the whole object, as shownin Figure 11. The differences of the mean cur-vature at each point is the criterion for con-structing the Scale Space. Figure 12 showsdetected keypoints of two differnt scans of thesame shoe.

Feature Descriptor

The feature descriptor transforms the area atthe detected keypoint into an easy compa-rable and meaningful description. Usuallythe approaches combine feature detectors andfeature descriptors into one method. Wellknown methods like MeshSIFT [16] or 3D-SURF [17] use their three-dimensional coun-terpart of feature descriptors developed fortwo-dimensional applications. Approachesusing Heat Kernel Signatures [19, 20] usethese for both – detection and description. Incontrast to that another approach called Spin-Images [21] is a feature descriptor only. It isable to describe an object locally or globally.The concept in [22] was adopted to a scale-invariant version encoding local information.Figure 13 shows a transformation of the areasurrounding keypoints into a 2D dense map


95

Figure 13: Transformation of shape features

using Spin Images [21]. Here a 3D mesh istransformed into several 2D maps, each re-lated to a keypoint

SO : R3 → R2

The 2D dense map is constructed using theequation

(α, β) = (√‖x− p‖2 − (n · (x− p))2, n·(x−p))

where (α, β) describe the new 2D coordinates.It is a cylindric coordinate system with itspoint of origin in the regarded point of themesh. A set of ranked Spin Images describesthe object itself, so it can be matched to theabstract brand model.Figure 14 summarizes the required steps forour shape matching algorithm using real worldobjects. The shape matching algorithm re-quires a three dimensional model of the prod-uct as input which can be matched to an ab-stract model of the brand product. The abstractmodel is a description of features that renderthe brand unique.

3.3 Odour Features

Much effort has been spent on how odourcould be measured. The European StandardEN-13725 [23] defines a method for the objec-tive determination of the odour concentrationof a gaseous sample using so called dynamicolfactometry. It is currently the only standard-ized method for the evaluation of odour im-pressions.The dynamic olfactometry is a method wherea panel of human assessors evaluates the con-centration of odour in a series of standard-ized presentations of a gas sample. Here the

Construct

Scale Space

Detect Key-Points

in Scale Space

User Input/

3D Model of Object

Shape Signature

Compute Local

Spin Images around

Key-Points

Compute Signature

from Spin Image Sets

Figure 14: Workflow for generating a shape signature

emission rate of odours emanating from pointsources, area sources with outward flow andarea sources without outward flow are consid-ered. The primary application of this standardis to provide a common basis for evaluationof odorant emissions in the member states ofthe European Union. Every method claimingthe ability to detect arbitrary odour emissionshas to benchmark against this standard. Anoverview of the development and applicationof electronic noses is given in Gardner andBartlett [24].In general it was observed that electronicnoses do not react to human inodorous gasesand were also unable to detect some gaseshumans are able to smell naturally. Begin-ning with the working principle of specific gas


96

Figure 15: Olfactory pattern of a genuine jersey (top)and a counterfeit (bottom)

sensors the concept of electronic noses as acombination of sensor array and diverse pat-tern recognition algorithms for classificationis introduced. In principle the sensor con-cepts could be divided into three categories.The commercially available electronic noseArtinos basing on the KAMINA (KArlsuherMIkroNase) [25] is a representative of metalconductance sensors. Here the sample gasflowing alongside the sensor surface is chang-ing the concentration and configuration of ox-ide containing compounds, thus changing theconductance of the metal-oxide, which is thenused as a measurement signal. The sensorelements differ by the thickness of silicondioxide coating. Additionally the temperatureis changed over time producing 38 analoguechannels containing also transient responses,which are to be analysed. Due to its workingprinciple these sensors deliver the most unspe-cific data, which is both an advantage and adisadvantage at the same time, since the sen-sors are suitable for a broad variety of samples,but the signal processing is harder to realise.A metal-oxide conductance sensor using 16channels was utilized in the project Inherent-ID [4].

A similar sensor setup is used in [26], thedifference being that the sensor elements are

coated with different polymers, which inducea change in conductance to specific gas com-ponents. It was shown that with four differ-ent sensor types held at four different temper-atures, so a total of 16 channels and follow-ing linear discriminant analysis ovarian can-cer could be detected from tissue samples.There are still some issues with falsely re-jected samples, but the results were quite im-pressive with respect to the use of ad-hocmethods. Another sensor concept utilisingpolymer coatings are the quartz microbalancesensor arrays as described in [27]. These sen-sors detect the change of frequency when agas is flowing over the sensor surface. Inprinciple these arrays are very sensitive butalso very susceptible to disturbances. Mostof recently published results in odour detec-tion are based on linear discriminant analy-sis and derivatives thereof. These methodsare efficient in classification of complex sen-sor data, but with a manageable number ofclasses. And these methods need a significantamount of data present and are therefore notsuitable for the here elaborated problem of oneto many matching, as needed for the applica-tion in counterfeit detection. An additional ob-stacle is the sensitivity to ambient conditionswhich result in wide variance of measurementdata from the same class of samples. Effortis made in the extraction of relevant featuresfor the purpose of reducing the dimensional-ity and the suppression of ambient influenceswhich was done by independent componentanalysis. An attempt of designing a generalodour model was made in [28], but was notsuccessful due to the sensors used and the factthat nonlinear behaviour was excluded in ad-vance. So the usage of specific models is morepromising.


97

Desired Signal Extraction

As it was described in the previous chapterthere are many ambient influences to odoursensing. For example humidity and temper-ature are different in Germany and Malaysia.Additionaly a mathematical expression forthe composition of odour is not linear, soodourous influences cannot be filtered out eas-ily. Given these facts and that the used ArtinosSensor returns most unspecific data it is a chal-lenge to filter environmental influences.

Figure 16: Independent Components of the measure-ments of a test object (left) and the environment (right)taken with an 16-channel multi gas sensor array, arb.unit

To meet the challenge of extracting desiredsignals in a robust fashion and filter the en-vironmental noise we use a similar approachto blind source separation, where two differentmeasurements are conducted. The first one is a

pattern from the environment without test ob-ject. The second is a pattern from the desiredsample in the beforementioned environment.The first signal can then be used to extract theplain odour of the object itself from the sec-ond signal. The components can be identi-fied and thus the ambient influence can be fil-tered. Since the electronic nose measurementdata delivers a nonlinear mixture of the envi-ronmental and sample odour there is no obvi-ous connection between these two patterns.One approach to divide the signals into theircomponents is the Independent ComponentAnalysis (ICA). Here the seperation is doneby statistical means. At most the ICA can re-turn as many independent components as thenumber of sensors used for capturing the inputdata, whilst reducing the complexity. In gen-eral the ICA has two major problems. The firstproblem is that the independent componentsare permuted. The sequence of two algorith-mic cycles might not be the same even with thesame data. The second problem is the loss ofvariance information in the independent com-ponents, since it cannot be restored.Figure 16 shows the independent componentsof a textile sample pattern on the left and theenvironmental reference pattern on the rightside. The independent components were ex-ctracted by an extended Bell & Sejnowski Al-gorithm [29] with adjusted break condition.Here the covariance criterion [30] was used.

E{g(u)uT} = I

If this equation is true the gi(yi) and yj are un-correlated for i 6= j. Therefore this can beseen as a nonlinear variant of principal com-ponent analysis.The next step after the ICA is to check the in-tegrity of the independent components. Thereare a some independent components whichseem to be noise. An autocorrelation analysisidentifies a possible noise contribution. Theseindependent components can be omitted. Af-


98

Figure 17: The Independent Components with reason-able high similarity measure are indicated by arrows.Noise contribution was omitted. arb. unit

terwards the similarity between the sampleand the environmental independent compo-nents are evaluated by applying the cosine dis-tance. The results are shown in Figure 17. Theindependent components with the strongestconnection are the independent componentswhich represent the environment in the sam-ple data and could be omitted as well. The ar-rows in Figure 17 indicate the correspondingindependent components.The exclusive signals of the sample measure-ment represent the core information on theodour of the test object. Figure 18 shows thedesired signal patterns which are exclusive forthe textile sample.

4 WORKFLOW

With the features described above there is astrong basis for automated classification ofpatterns. The key point for a robust and re-liable counterfeit detection is the combinationof these features and additional user informa-tion with the aim to derive a decision wetherthe probe is likely to be a counterfeit. An ad-vantage of the proposed algorithms for feature

Figure 18: Core information of a textile sample, arb.unit

UserInput

2D ImageFeatures

Odour Features

3D Features

Classificator/Decission

Figure 19: Concept of feature fusion

extraction is the possibility to utilize statisticalframeworks since the features are representedby probability density distributions.In general there are various approaches possi-ble. Starting with a direct fusion of the fea-tures as proposed in [31] and shown in Figure19, or a more sophisticated approach whichis taking the process of probing into account.Such a workflow is depicted in Figure 20.Here the decision process is not necessarilybased on the utilization of all features, sincesome of them are dispensable or could be mis-leading. Think of the probing of shirt, obvi-ously the 3D geometry cannot give a relevantcontribution to the decision process and the


99

UserInput

Logo Matching Textur Analysis

3D Matching

Classificator/Decission

2D ImageCapturing

Capture3D Model

Acquire3D Data?

AcquireSmell Data?

N

Y

CaptureOdour

Odour Matching

N

Y

Figure 20: Sophisticated workflow for counterfeit de-tection

3D scanning can therefore be ommitted. Theclassification itself is done with an adjustedBayesian approach where special account wasgiven to the detection of novel and thereforeunknown patterns. This was done with estima-tion of the Level of Signifcance distribution,which gives a decision information and an ad-ditional value of the plausibilty of this deci-sion, cf. [32].

5 CONCLUSION

It was shown that the Inherent-ID Projectadopts a novel approach to protecting high-value products from counterfeiting. The ap-

Figure 21: Future Scenario for Counterfeit Detection

proach is based on the stationary and mobilecapture of key product features indissolublylinked with the product which enable its pro-duction process to be traced. This not onlyrenders the application of security tags obso-lete but also gives enhanced protection againstcounterfeiting as the inherent characteristicsthat the high-quality production process im-pregnate in the genuine product are combinedwith one another to serve as proof of prod-uct identity. They form the basis on whichelectronic certificates of authenticity can be is-sued without the need for complicated explicitsecurity markings. Methods for the captureand control of identity characteristics are be-ing elaborated in the Inherent-ID project forsystem integration using intelligent camerasand an electronic nose. The identity character-istics captured by this range of sensors serveboth for the product identification and prod-uct authentication. At the same time this alsooffers opportunities for improving documenta-tion of product flows in the supply chain. Fulldocumentation serves as a complement to theinherent characteristics of the authentic prod-uct and offers valuable information of verifi-cation of the genuine article, thus serving tosafeguard against counterfeits.

6 PERSPECTIVE

The approach of the project Inherent ID can beadopted to a possible future scenario for coun-terfeit detection. As shown in Figure 21 theapproach could be ported to work with con-


100

sumer electronics like smartphones, since 3Dcameras are already available there. The textu-ral features and the shape features of an objectcould be detected with the built-in cameras.The classification itself can then be done withan approch using Service Orientend Archi-tectures (SOA), where the features are trans-fered from the smartphone over the internet toa server. This is necessary because even re-cent smartphones with multicore cpu’s are tooslow to compute the proposed algorithms in atimely fashion.This enables not only customs officials to de-tect counterfeits, any customer would be ableto do that using the detection app. This couldlead to a whole new market driven combatagainst product piracy.

7 ACKNOWLEDGEMENTS

The authors would like to acknowledge thefunding of the research project Inherent-ID bythe senate of the state Berlin and the EuropeanRegional Development Fund. The project isembedded in the Fraunhofer Cluster of In-novation Secure Identity Berlin Brandenburg.Furthermore we would like to acknowledgethe work done by our students Evelyn Jung-nickel and Norman Franke in the project.

8 THE AUTHORS

The authors are working within the Automa-tion Group at Production Technology Center(PTC) Berlin, Germany. The PTC comprisesthe department of Industrial Automation Tech-nology at Technische Universitat Berlin andthe Fraunhofer Institute for Production Sys-tems and Design Technology.The main tasks of the Automation Group arefundamental research and lecturing in a broadband of topics regarding industrial automationsuch as process automation and robotics, pro-

cess monitoring and simulation, image pro-cessing and pattern recognition.

9 REFERENCES

[1] European Comission, Report on EU customs en-forcement of intellectual property rights – Re-sults at the EU border - 2011, ISBN 978-92-79-25362-1, pp. 10-19, 2012

[2] OECD, The Economic Impact of Counterfeitingand Piracy, OECD, Paris, 2008, www.oecd.org/sti/counterfeiting

[3] M. Abramovici, L. Overmeyer, B. Wirnitzer,Kennzeichnungstechnologien zum wirksamenSchutz gegen Produktpiraterie - Band 2 derReihe Innovationen gegen Produktpiraterie,ISBN 978-3-8163-0602-3, pp. 1-196, 2010

[4] J. Kruger, M. Blankenburg Secure Identity - asource for innovative it-systems and processesFuture Security, 5th Security Research Confer-ence, Berlin, September 7th - 9th, 2010. Ed.:Fraunhofer VVS, Freiburg, 2010

[5] M. Tuceryan, A. K. Jain, Texture Analysis,Handbook of Pattern Recognition & ComputerVision, 2nd Edition, World Scientifc PublishingCo. Ptc. Ltd., 2001

[6] Y. Xu, H. Ji, C. Fermuller, Viewpoint InvariantTexture Description Using Fractal Analysis, IntJ Comput Vision 83, pp. 85-100, 2009

[7] M. Varma, A. Zisserman, Classifying images ofmaterials: Achieving viewpoint and illuminationindependence, ECCV Volume 3, pp. 255-271,2002

[8] M. Varma, A. Zisserman, Texture Classification:are filter banks necessary?, CPVR Volume 2,pp. 691-698, 2003

[9] B. Jahne, Digital Image Processing, ISBN 3-540-24035-7, 2005

[10] Besl P, and McKay N. A Method for Registrationof 3-D Shapes. IEEE Transactions on PatternAnalysis and Machine Intelligence, 14(2):239-256. 1992


101

[11] J. W. Tangelder, R. C. Veltkamp A Survey ofContent Based 3D Shape Retrieval Methods,Multimedia Tools and Applications Vol. 39 No.3. , pp. 441-471, 2007

[12] P. Shilane, P. Min, M. Kazhdan, T. FunkhouserThe princeton shape benchmark, Shape Model-ing International, 2004

[13] A. M. Bronstein et al.SHREC 2010, Proc. EU-ROGRAPHICS Workshop on 3D Object Re-trieval (3DOR), 2010

[14] Sipiran I and Bustos B. A Robust 3D Inter-est Points Detector Based on Harris Operator.Proc. EUROGRAPHICS Workshop on 3D Ob-ject Retrieval (3DOR), 2010.

[15] Flitton G, Breckon T, and Bouallagu N. ObjectRecognition using 3D SIFT in Complex CT Vol-umes. In Proc. British Machine Vision Confer-ence, pages 11.1-11.12. BMVA Press, 2010.

[16] Maes C., Fabry T., Keustermans J., Smeets D.,Suetens P., Vandermeulen D. Feature detectionon 3D face surfaces for pose normalisation andrecognition. In Proc. BTAS, 2010.

[17] Knoop J et al. Hough Transform and 3D SURFfor robustthree dimensional classification. InProc ECCV. 2010.

[18] Litman R, Bronstein A, Bronstein M. Diffusion-geometric maximally stable component de-tection in deformable shapes. Arxiv preprintarXiv:1012.3951. 2010.

[19] Sun J, Ovsjanikov M, Guibas L. A conciseand provably informative multi-scale signaturebased on heat diffusion. Eurographics Sympo-sium on Geometry Processing (SGP). 2009.

[20] Bronstein A, Bronstein M, Guibas L and Ovs-janikov M. Shape google: Geometric words andexpressions for invariant shape retrieval. ACMTransactions on Graphics (TOG) Volume 30 Is-sue 1. 2011.

[21] A. Johnson, M. Hebert, Using spin images for ef-ficient object recognition in cluttered 3d scenes,IEEE PAMI 21, pp. 433-449, 1999

[22] Darom T, Keller Y. Scale invariant features for3dmesh models. IEEE Transactions on ImageProcessing. 2010.

[23] EN 13725, Air quality. Determination of odourconcentration by dynamic olfactometry, DINEN 13725:2003

[24] J. W. Gardner, P. N. Bartlett, Electronic Noses -Principles and Applications, Measurement Sci-ence and Technology Volume 11, 1999. OxfordUniversity Press, Oxford

[25] D. Haeringer, J. Goschnick, Characterization ofsmelling contaminations on textiles using a gra-dient microarray as an electronic nose, Sensorsand actuators B-Chemical Vol. 132, Nr. 2, 2008

[26] J. Chilo, G. Horvath, T. Lindblad, R. Olsson,Electronic Nose Ovarian Carcinoma DiagnosisBased on Machine Learning Algorithms, Lec-ture Notes in Computer Science Volume 5633,2009

[27] A. S. Yuwono, T. Hamacher, J. Nieß, P. Boeker,P. S. Lammers, Implementation of a quartz mi-crobalance (QMB) sensor array - based instru-ment and olfactometer for monitoring the per-formance of an odour biofilter, 2nd IWA Inter-national Workshop & Conference on Odour &VOC’s, Singapore, 2003

[28] F. Bitter, Modell zur Bestimmung der Geruchsin-tensitat der Raumluft mit Multigassensorsyste-men, doctoral thesis, TU Berlin, 2009

[29] Bell, A.J., Sejnowski, T.J. An information max-imisation approach to blind separation andblind deconvolution, Neural Computation, 7, 6,pp. 1129-1159, 1994

[30] Hyvaerinen, A., Karhunen, J., Oja, E. Inde-pendent Component Analysis,ISBN 978-0-471-46419-8 John Wiley and Sons, pp. 229-232,2004

[31] H. B. Mitchell Multi-Sensor Data Fusion: AnIntroduction, Springer publishing, 2007

[32] S. Kuhn, Stochastic Engineering – Berechnung,Entwicklung und Modellierung bei unsichererInformation, doctoral thesis, TU Berlin, ISBN978-3-8322-9188-4, pp. 103-115, 2010


102

product piracy prevention: product counterfeit detection...

Documents