High volume security printing using sheet-fed offset press Slavtcho (Slavi) Bonev Epyxs GmbH Richard-Wagner-Str 29, 6816 Mannheim, Germany sbonev@epyxscom Abstract: Security printing based on DataGrid and EpiCode is a novel, very efficient anti-counterfeiting technique based on the individuality of prints DataGrid is a high-density halftone data storage which forms a unique signature the so called EpiCode when it is printed The EpiCode is caused by the physical interaction between printing medium (ink) and substrate (paper) and is used as a fingerprint for authentication In sheet-fed offset printing the signature of the printing plate is a novel feature for anti-counterfeit applications In the presented printing tests, the influence of the substrate on the EpiCode and the long-term stability of the offset printing process regarding the plate signature are determined using biometrical analysis Equal error rates (EER) better than -4 are achieved in the case of four paper grades commonly used for labels and packaging Further, the printing plate signature has proven a stabile feature for a cluster of printed DataGrids and is therefore called ClusterCode Authentication based on ClusterCode achieved equal error rates (EER) better than - In addition, a recognition system using a reduced database is proposed Keywords: security printing, paper substrate, long-term stability, print quality control, biometrical system 1 Introduction Each EpiCode is uniquely linked to a printed DataGrid (Figure 1) This individual print signature (Bon, 28) comprises at least two independent components: (i) a stochastic component from the medium-substrate interaction eg microscopic ink smudging on the paper surface and (ii) a stochastic component from the irregularities of the printing plate It is expected that the surface roughness will have an impact on the first component - a) 1 2 Figure 1: Matrix code DataGrid (a) composed of 14x14 graphic symbols and its corresponding fingerprint signal EpiCode (b) of 196 samples with an amplitude resolution of 4 bits per sample The printing plate will have a systematic impact on the EpiCodes in a cluster A cluster is a batch of DataGrids on consecutive printed sheets which originates from the same master DataGrid on the printing plate So the second stochastic component, which presents the printing plate signature, is called ClusterCode and is comprehensively analysed here Marking product packaging and labels with DataGrid is a very effective low-cost approach for anti-counterfeit applications for mass products The DataGrid is simply inserted in the package/label design before printing and carries encrypted information about the protected product and a pointer to a database In addition the EpiCode is extracted from each printed DataGrid and is used to identify the corresponding item 2 Methods Fingerprint identification is based on two basic premises: (i) persistence: the basic characteristics of fingerprints do not change with time; and (ii) individuality: the fingerprint is unique to an individual (Pan, 22) The same assumptions are valid for an iris recognition system using the IrisCode (Dau, 1999) Further, (iii) a reasonable alignment between a template (stored in the database) and the input (which needs to be identified) should be established to ensure a reliable comparison To identify the genuine, which means to find the best match between input and template, (iv) an exhaustive search in a large data base is necessary b) iarigai, Advances in Printing and Media Technology, Volume XXXVI, No 4, pp 449-4, ISBN 978-3-981274-1-9 Page 1
The technology presented in this paper goes beyond standard fingerprint applications in two aspects Since the EpiCode is adherent to the data, an exact alignment is achieved by the image preprocessing, where the code symbols are detected and used as a coordinate system to realize a repeatable measurement Secondly, using the process specific property of sheet-fed offset printing, eg the systematic influence of the printing plate on the EpiCode, the search procedure for identification (one-to-many comparison) could be reduced to a verification (one-to-one comparison) To analyze the influence of the paper roughness on the EpiCode, four different grades of paper (Figure 2) and 3 sheets of each paper grade were printed in the same print run The printing process parameters and the measurement settings, including the used devices and materials, are described in detail in Table 1 Figure 2: Microscopic images (8x) of the four paper substrates (s Table 1) made with the KEYENCE VHX-6 microscope The surface structures are evidently different For example, paper Nr 2 (b) is a glossy cast-coated paper A single printed code symbol measures 127 x 127 µm (height x width) Device/Material Model Comments Platesetter AGFA Xcalibur VLF 24 dpi physical printing resolution, Computer-to-Plate Printing plate AGFA Thermostar P97 83 nm sensitive plate Offset press MAN ROLAND 6 LV Six colours offset press Printing ink Printcom S112Y Black fast drying ink Coating Printcom S24U Dispersion coating, water based varnish Printing unit - Last printing unit, no overrolling, no drying Matrix code DataGrid 3 x 3 graphic symbols, 38 x 38 mm (height x width) Flatbed scanner HP scanjet 82 24 dpi scan resolution is used (48 dpi optical resolution) Paper grades (Substrates) Nr 1: Twin Coat Nr 2: ZANDERS Mega gloss Nr 3: Niklaselect (Bekk 12) Nr 4: Niklakett Brilliant (4) Long grain, x 7 cm, 3 g/m 2, 46 mm, 3 sheets Long grain, 2 x 74 cm, 1 g/m 2, 13 mm, 3 sheets Short grain, 2 x 74 cm, 8 g/m 2, 7 mm, 3 sheets Short grain, 2 x 74 cm, 8 g/m 2, 7 mm, 3 sheets Table 1: Detailed description of the printing process and the measurement conditions To estimate the discriminatory power of the EpiCode a biometrical analysis of the results is performed (Bon, 28) As a result of a comparisson of two different EpiCodes, one from the reference set (database) and one from the authentication set (probe), the cross is evaluated The higher the coefficient, the higher is the similarity between both EpiCodes All possible comparisons are performed and the correlation coefficients are depicted in a histogram Gaussian distributions are fitted to the measurements and (false acceptance rate) and FRR (false rejection rate) are calculated for defined thresholds of discrimination Finally, the EER (equal error rate) achieved is presented in a ROC (reciever operating characteristic) diagram iarigai, Advances in Printing and Media Technology, Volume XXXVI, No 4, pp 449-4, ISBN 978-3-981274-1-9 Page 2
reference scan enrollment set authentication scan authentication set n n s s c c Figure 3: Experimental concept: n code number (N=8), c code info (C=8), s sheet number (S=8) Short description of the constellation of comparisons in relation to the color coding is given below: yellow: codes with different information different plate region resp location on sheet (false matches) blue: codes with identical information different plate region resp location on sheet (copies) green: codes with identical information same plate region resp location on sheet (copies in cluster) red: same code identical information, same plate region resp location on sheet (originals) 3 Results Two paper grades, Twin Coat and ZANDERS Mega gloss, are analysed and the results are presented in Figure 4 The copies in a cluster, printed by the same region of the printing plate, show high correlation values This subset of copies contains obviously the printing plate signature which causes a systematic influence on the fingerprint signal (of the subset) called ClusterCode Since only the paper substrate is changing during the printing, the influence of the paper roughness can also be validated The stochastic influence of the substrate surface will try to reduce the Therefore the glossy paper shows higher mean value of the cluster distribution The recognition performances (EER) are estimated regarding EpiCode (print signature, red colored distribution) and ClusterCode (plate signature, green colored distibution) and are presented in Figure As a reference, the distribution of results for printed copies (blue colored) is considered Prob density 1 Paper Nr 1 EpiCode: THR = 467, EER = 1e-48 ClusterCode: THR = 264, EER = 1e-11 µ = 7, σ = 34 µ = 77, σ = 37 µ = 48, σ = 44 µ = 71, σ = 27-1 1 2 3 4 6 7 8 9 1 11 1 Paper Nr 2 EpiCode: THR = 417, EER = 1e-36 ClusterCode: THR = 289, EER = 1e-14 a) Prob density µ = 7, σ = 36 µ = 79, σ = 37 µ = 61, σ = 49 µ = 72, σ = 34-1 1 2 3 4 6 7 8 9 1 11 Figure 4: Results of the measurements of 8 consecutive sheets (12 EpiCodes) for paper substrates Twin Coat (a) and ZANDERS Mega gloss (b) A significant differencebetween both ClusterCodes can be observed The EpiCode is a 9-element vector with an amplitude resolution of 4 bits per element b) iarigai, Advances in Printing and Media Technology, Volume XXXVI, No 4, pp 449-4, ISBN 978-3-981274-1-9 Page 3
Using high-quality printing processes, as the experiment shows, recognition performances of around -4 using EpiCode and around - using ClusterCode are practically achieved ROC curves for EpiCode ROC curves for ClusterCode -2 - FRR -4-6 -8 - - -8 Paper Nr 1: EER = 1e-48 Paper Nr 2: EER = 1e-36-6 -4-2 - -1-2 -2-3 -3-2 Paper Nr 1: EER = 1e-11 Paper Nr 2: EER = 1e-14 Figure : Estimated EER (equal error rates) using EpiCode (print signature) and ClusterCode (plate signature) The amplitude resolution of the EpiCode is reduced from 4 bits/sample (optimum) to 1 bit/sample (minimum) to simulate worst-case conditions caused eg by defocused scan and to estimate the minimal recognition performance (Table 2) This is also important if the size of the database has to be reduced to the minimum for a certain reason, for instance, improvement of storage space or search time The corresponding EER decrease from ~ -4 to ~ -2 for EpiCode and from ~ - to ~ -6 for ClusterCode -2-1 - - Threshold / EER Threshold / EER Threshold / EER Threshold / EER Nr Paper grade 4-bit EpiCode 1-bit EpiCode 4-bit ClusterCode 1-bit ClusterCode 1 Twin Coat 467 / 1e-48 31 / 1e-2 264 / 1e-11 183 / 1e-6 2 ZANDERS Mega gloss 417 / 1e-46 28 / 1e-19 289 / 1e-14 22 / 1e-8 3 Niklaselect 484 / 1e-9 319 / 1e-28 261 / 1e-13 18 / 1e-7 4 Niklakett Brilliant 46 / 1e-48 36 / 1e-2 26 / 1e-12 174 / 1e-7 Table 2: Estimated recognition performance using the printing process described in Table 1 The paper grade with the best behaviour regarding the ClusterCode (Figure 4b) is selected for the analysis of the long-term stability of the printing process and its impact on the recognition performance A stable process would produce a constant ClusterCode over the whole printing cycle For each thousand sheets a probe of four sheets is analysed (Figure 6) The cross s of all EpiCodes are compared; the distributions of results were taken into account The measurement confirms the stability of the plate signature over 3 printed sheets for the high-quality printing process An example of a low-quality printing process is given below 1 1 21 31 Figure 6: Constellation of the sheet sampling in the long-term stability test: A stack of 3 sheets is printed After each prints 4 sheets are extracted to measure the long-term stability of the ClusterCode Though the ClusterCode remains stable over 3 sheets, certain decrease in EER can be observed (Figure 7) Due to slight changes of some parameters during the printing, deviations in the plate fingerprint signal occur This is a slow process which has no impact on the ClusterCode measured on consecutive sheets iarigai, Advances in Printing and Media Technology, Volume XXXVI, No 4, pp 449-4, ISBN 978-3-981274-1-9 Page 4
Prob density 1 Paper Nr 2 EpiCode: THR = 417, EER = 1e-36 ClusterCode: THR = 289, EER = 1e-14 µ = 7, σ = 36 µ = 79, σ = 37 µ = 61, σ = 49 µ = 72, σ = 34-1 1 2 3 4 6 7 8 9 1 11 1 Paper Nr 2 EpiCode: THR = 461, EER = 1e-41 ClusterCode: THR = 26, EER = 1e-8 a) Prob density µ = 13, σ = 3 µ = 99, σ = 38 µ =, σ = 61 µ = 731, σ = 28-1 1 2 3 4 6 7 8 9 1 11 Figure 7: Long-term stability test for sheet-fed offset print on the substrate ZANDERS Mega gloss (Paper Nr 2): Comparison of the measurements on 8 consecutive sheets (a, 12 EpiCodes) and 3 sheets (b, 24 EpiCodes) A slight offset and widening of the distribution of results for ClusterCode (b) can be observed due to process deviations over 3 sheets The ClusterCode remains stable, the EER decreases from -14 to -8 To understand the difference between low-quality and high-quality printing processes regarding security applications, the behaviour of a quality parameter is shown in Figure 7 The quality parameter is defined according to the theory of digital communications and is a measure of the code symbol transmission quality Its value is calculated before decoding and can also be used directly for print quality control For critical values of the quality parameter below 2 the printed code is considered as non-readable and an automatic identification of the DataGrid is impossible The high-quality printing process is realized in the experimental print run described in Table 1 The low-quality process is realized in an industrial print run using standard paper and ink At each thousand printed sheets one is extracted for control measurement Nearly 2 sheets are printed with 12 DataGrids per sheet (1 sheets measured), the high-quality process is represented with 64 DataGrids per sheet and only 3 sheets (4 sheets measured) The distribution of results for the copies in a cluster, which is not presented here, shows extremely high standard deviation for the low-quality process Generally, significant differences in mean value and standard deviation can be observed According to this, the low-quality process will produce non-readable codes especially at the end of the print run, where the standard deviation increases visibly This clearly shows, that not all printing processes are capable for security printing designed for the usage of automatic authentication algotrithms b) quality parameter 4 4 3 3 2 2 quality variation during production high quality process low quality process 1 1 2 3 4 6 7 8 9 11 12 13 14 1 16 17 18 19 2 21 22 sheet number / Figure 7: Comparison between estimated values of the quality paramater for high quality and low quality printing processes The error plots show a mean value / standard deviation of 39 / 1 and 26 / 2 for the high and low quality processes, respectively DataGrid codes with a quality parameter lower than 2 cannot be read due to failures in the forward error correction (FEC) algorithm iarigai, Advances in Printing and Media Technology, Volume XXXVI, No 4, pp 449-4, ISBN 978-3-981274-1-9 Page
A strong decrease of the recognition performance for the low-quality process is confirmed for both recognition approaches (Figure 8) Due to the small EER for ClusterCode and the produced defect DataGrids, security printing using low-quality or non-approved printing processes is not recommended ROC curves for EpiCode ROC curves for ClusterCode -2 - FRR -4 - -6-8 HQ process: EER = 1e-41 LQ process: EER = 1e-22-8 -6-4 -2-1 -2 HQ process: EER = 1e-8 LQ process: EER = 1e-1-2 -1 - - Figure 8: Comparison between estimated EER for high quality (HQ) and low quality (LQ) printing processes 4 Discussion The quantitative comparison of the individual code properties shows a clear dependance between paper surface structure and the achieved EER Rough paper surface will have negative impact on the ClusterCode and a low impact on the EpiCode For a glossy paper printed with fast drying ink it could be shown, that the ClusterCode remains stable during thousands of prints when a high-quality process is used Nevertheless using the introduced quality parameter, quality control can be performed on item level, also by using images from conventional print quality control devices The long-term stability of the ClusterCode could be confirmed A further investigation of the long-term behaviour over a complete production run must be done since it is known that properly developed (post-baked) printing plates may exceed run lengths of a million impressions (Wik, 28) The reliable detection of the plate fingerprint in printed DataGrid codes (ClusterCode approach) can be used for a new kind of security applications based on identification of original prints using reduced databases as shown in Figure 9 Using scans of all printed copies on a single sheet, all product packages produced in a print run can be identified Scanning more sheets will lead to an increase of the size of the database, but also to an increase of the authentication performance of the recognition approach sheet nr 1 High reliability: EER ~ -41-36 based on EpiCode (all sheets scanned) reference scan of each single package large database, 24 entries sheet nr 21 sheet nr 31 Middle reliability: EER ~ -2 (approx) based on ClusterCode (multiple sheets) reference scan of 2 sheets small database, 24 entries Low reliability: EER ~ -8 based on ClusterCode (single sheet) reference scan of 1 sheet very small database, 12 entries Figure 9: Principal of multilevel-security printing, based on DataGrid, for a standard print run with 2 printed sheets and 12 copies per sheet (24 product packages) iarigai, Advances in Printing and Media Technology, Volume XXXVI, No 4, pp 449-4, ISBN 978-3-981274-1-9 Page 6
Conclussions and future prospects Using the unique printing plate signature ClusterCode the feasibility of a novel security application has been shown Printing processes with a long-term stability regarding to the ClusterCode should be identified by appropriate measurements Industrial printing field tests will include the use of several sheet-fed offset presses with (high-durability) post-baked printing plates Recently, new approaches for a simplified database search are implemented and tested Further developments are running to combine the advantages of standard mobile coding technologies with the multilevel security options of the DataGrid code (Mal, 29) An example is given in Figure by bicolor DataGrids with DataMatrix code for mobile applications implemented as a color overlay color overlay code image preprocessing DataMatrix low data density automatic color separation basic code DataGrid high data density digital filtering fingerprint EpiCode Figure : Bicolor DataGrids in yellow-black and cyan-magenta with implemented DataMatrix code Acknowledgements We gratefully acknowledge the financial support by the Federal Ministry of Education and Research of Germany (FKZ 2PU11, OPUR project, wwwopur-securede) We acknowledge the support of Mr Rainer Gebhardt and Mr Markus Jung-Diefenbach from manroland AG for performing the high-quality printing tests and providing specific information about the research Many thanks to Mr Thomas Buss from GEWA Etiketten GmbH for supervising the industrial field test References [1] Bonev S and Wirnitzer B, 28, Security printing for product packaging in industrial printing applications, Advances in Printing and Media Technology, 3th International Research Conference of IARIGAI [2] Pankanti S, Prabhakar S and Jain A K, 22, On the individuality of fingerprints, IEEE Transactions on Pattern Analysis and Machine Intelligence, 24, 8, -2 [3] Daugman J, 1999 Recognizing Persons by Their Iris Patterns, in Biometrics: Personal Identification in Networked Society, A K Jain, R Bolle, and S Pankanti, editors, Kluwer Academic Publishers [4] Wikipedia, 28, http://enwikipediaorg/wiki/offset_printing forensic data [] Maleshliyski S, García F: Integration of anti-counterfeiting features into conventional 2D barcodes for mobile tagging, 29, TAGA 61 st Technical Conference, 1-18 March, 29, New Orleans, USA iarigai, Advances in Printing and Media Technology, Volume XXXVI, No 4, pp 449-4, ISBN 978-3-981274-1-9 Page 7