Signature-Embedding In Printed Documents For Security and Forensic Applications

Signature-Ebedding In Printed Docuents For Security and Forensic Applications Aravind K. Miilineni a, Gazi N. Ali a, Pei-Ju Chiang b George T. C. Chiu b, Jan P. Allebach a, Edward J. Delp a a School of Electrical and Coputer Engineering b School of Mechanical Engineering Purdue University, West Lafayette, Indiana USA ABSTRACT Despite the increase in eail and other fors of digital counication, the use of printed docuents continues to increase every year. Many types of printed docuents need to be secure or traceable to the printer that was used to print the. Exaples of these include identity docuents (e.g. passports) and docuents used to coit a crie. Traditional protection ethods such as special ins, security threads, or hologras, can be cost prohibitive. The goals of our wor are to securely print and trace docuents on low cost consuer printers such as injet and electrophotographic (laser) printers. We will accoplish this through the use of intrinsic and extrinsic features obtained fro odelling the printing process. Specifically we show that the banding artifact in the EP print process can be viewed as an intrinsic feature of the printer used to identify both the odel and ae of the device. Methods for easuring and extracting the banding signals fro docuents are presented. The use of banding as an extrinsic feature is also explored. Keywords: docuent security, secure printing, printer identification, banding. INTRODUCTION In today s digital world securing different fors of content is very iportant in ters of protecting copyright and verifying authenticity. Many ethods 7 have been developed to protect audio, video, digital docuents, iages, and progras (executable code). In this paper, we will be concerned with the securing of printed docuents, which we will refer to as docuents. The securing of docuents is not new. In 282 AD physical paper waterars began to appear. 8 These were created by placing thin wires in the old used to create the paper. The purpose of these ars is to identify the paper aer or the specific old used to create the paper. By the early eighteenth century these types of waterars started to appear in ban notes to deter counterfeiting and also on regular parchent to designate a tradear or a particular date or place. Today, paper wateraring is still used as can be seen in ost ban notes and soe official docuents, but it is usually accopanied by other security features. Security fibers or threads can be ebedded within the paper or woven into it during anufacture. Special ins which change color or react to certain cheicals which ight be used to rear a docuent have been proposed. 9 Additionally, new ethods using ebedded hologras, or icrotext are being introduced. These ethods are used to verify the authenticity of a docuent and usually do not carry any other inforation. For ore detail on these types of security features see 2. In digital data hiding, any inforation (a waterar), such as copyright and owner inforation, can be ebedded within a edia data eleent such as an iage or video sequence. 4 The waterar could be visible or non-visible, but in either case should be easily recoverable fro the waterared data. 5 Additionally the waterar ust be able to survive if the host data is anipulated., 3 We believe that a aring schee analogous to digital wateraring but for docuents is very iportant. Printed aterial is a direct accessory to any criinal and terrorist acts. Exaples include forgery or alteration This research was supported by a grant fro the National Science Foundation, under Award Nuber 29893. Address all correspondence to E. J. Delp at ace@ecn.purdue.edu

of docuents used for purposes of identity, security, or recording transactions. In addition, printed aterial ay be used in the course of conducting illicit or terrorist activities. Exaples include instruction anuals, tea rosters, eeting notes, and correspondence. In both cases, the ability to identify the device or type of device used to print the aterial in question would provide a valuable aid for law enforceent and intelligence agencies. We also believe that average users need to be able to print secure docuents, for exaple boarding passes and ban transactions. The proble is that the use of special papers, special ins, or hologras can be cost prohibitive. Most of these techniques either require special equipent to ebed the security features, or are siply too expensive for an average consuer. Additionally, there are a nuber of applications in which it is desirable to be able to identify the technology, anufacturer, odel, or even specific unit that was used to print a given docuent. We propose to develop two strategies for printer identification based on exaining a printed docuent. The first strategy is passive. It involves characterizing the printer by finding intrinsic features in the printed docuent that are characteristic of that particular printer, odel, or anufacturer s products. We shall refer to this as the intrinsic signature. The intrinsic signature requires an understanding and odelling of the printer echanis, and the developent of analysis tools for the detection of the signature in a printed page with arbitrary content. The second strategy is active. We ebed an extrinsic signature in a printed page. This signature is generated by odulating the process paraeters in the printer echanis to encode identifying inforation such as the printer serial nuber and date of printing. To detect the extrinsic signature we use the tools developed for intrinsic signature detection. We will use our nowledge of the printer echanis and the results of the printer characterization to deterine the printer process paraeters that can be odulated to encode the desired identifying inforation. The odulation of these paraeters will require odification of the printer echanis. 2.. Hiding Data in Text 2. OTHER APPROACHES Various ethods for securing docuents have been developed. One of the earliest ethods for securing printed text involves the shifting of eleents in a text docuent as described in 4 6. In this approach, ethods are described for providing copyright protection for text docuents by ebedding inforation specific to the recipient or source in order to deter illegal distribution. The ethods developed to encode this inforation into each page involve the shifting of textual eleents in the docuent by aounts iperceptible to the huan viewer to encode individual bits of data. These textual eleents can be lines, words, or individual characters. With line shifting, every other line is shifted slightly up or down, approxiately /6th inch, to encode a one or zero. To detect the shifts in a docuent, no prior inforation about the original is necessary since the inforation is ebedded by shifting every other line. The baseline locations can be estiated in the scanned docuent and by easuring the relative distances between the baselines of adjacent text lines an estiate of the ebedded data can be found. It is shown that this type of encoding is robust to scan-print attacs and photocopy generation loss, although soe variability in the detection arises due to errors in the scan process such as rotation. Another detection ethod for line shifting, ore robust to iaging errors such as scan rotation, is the use of the relative distance between centroids of adjacent lines of text. For this detection ethod the original docuent is needed in order to properly decode the inforation because the centroids are dependent on the content on each text line. Word and character coding allow ore data to be ebedded into a page of text, but are not as robust as line coding due to the fact that each shift is encoded in a saller portion of the printed page. Also ost word processors will vary the spacing between adjacent words and characters, so estiation of the shifts using the differential ethod used for lines will not wor in this case unless the original docuent is also available. These shifts could instead be used as a fragile waterar to detect alterations to a docuent. One ajor drawbac to these feature shifting ethods is that they are easily defeated. An attacer can siply scan the docuent and use readily available tools to extract and reforat the text, reoving any shifts, and reprint it without the encoded inforation.

2.2. Hiding Data with Halftone Patterns The previous ethod deals with encoding inforation in text, but docuents ay also have halftone iages. Halftoning is the process of converting a continuous tone iage into an iage having only a finite nuber of levels, typically 2 for printed iages. 7 When viewed fro a distance, these iages reseble the original. Nuerous ethods have been developed for wateraring halftone iages. Most of these ethods involve odifying the halftone patterns used when printing the iage. The three basic ethods are the use of ultiple dithering atrices, conjugate halftone screens, or angular variation of the halftone screen. 8 In the first ethod, the dithering atrix changes fro tile to tile and can be used to encode inforation. Detection involves easuring the statistical properties of the halftone patterns and their variation across the iage. The second ethod involves conjugate halftone screens, two screens are used to for two halftone iages and the data are ebedded through the correlations between two screens. 8 The third basic ethod involves encoding the data in the angular orientation of the screen in each halftone cell. In this case, each angular orientation can represent ultiple bits depending on the nuber of angles the halftone screen can be generated and detected at. Another recent ethod of data hiding with halftone patterns has been described in 9 in which waterars are ebedded in duplex printed pages such that when the page is held up to a light source a fae waterar will appear. This ethod relies on halftone iages that are printed opposite each other on either side of the page. This technique requires a high degree of control over the registration of each side of the docuent to ae sure the halftone patterns line up. A siilar technique is also described in 2 which can be used for one sided printing with verification using a transparent screen. 2.3. Data Hiding with Wateraring Techniques for Continuous Tone Iages 4, 5, 7 Wateraring iages in docuents can also be done using continuous tone iage wateraring techniques. These ethods first ebed a waterar into the continuous tone iage and then print it at a high resolution to create the docuent. To detect the waterar, the docuent is scanned and transfored bac into a continuous tone iage after which an appropriate ethod for detecting the waterar is used. The type of waterar ebedded has to be one that can survive the print-scan process when viewed as an attac channel. The wateraring ethod described in 2 23 has been shown to be very robust against any types of attacs, including the print-scan attac and those involving global affine or local non-linear transforations. Being able to survive these two types of transforations is iportant because a docuent ight be wrinled or torn. First a copressed and encrypted essage is encoded into a reference waterar creating a sall bloc. The bloc is then flipped and irrored to create a acrobloc consisting of 4 copies of the original bloc. The purpose of the flipping is to visually decorrelate the ebedding structure and also to assist in later estiation of local non-linear transforations. The acrobloc is then tiled and ebedded into the iage which is then printed. An autocorrelation detection schee as described in 24 can be used to estiate any global affine transforations that have been applied to the scanned docuent. Local non-linear transforations can then be estiated using a siilar ethod within each acrobloc. Until recently, none of the ethods which waterar continuous tone iages before printing have explicitly taen into account the printing process itself. The ethod described in 25 jointly optiizes the printed iage quality and the detectability of the waterar. This approach uses direct binary search (DBS) halftoning 26 and a spread spectru waterar. 27 The halftoned iage is odeled as a bitap. Because the detection process relies on a reconstructed continuous tone version of the docuent, and not the specific halftone patterns used in its creation, there are ultiple bitaps of the sae waterared iage which will yield good detection results. Assuing that the bitap iage is M N pixels, there are then 2 MN possible bitaps, only a subset of which will visually reseble the original iage data and siultaneously allow detection of the waterar. Each of these bitaps will differ in visual quality and the goal is to pic the one which best axiizes both visual quality and waterar detectability. This is achieved by using a odified version of DBS such that at each iteration both the waterar detectability and perceptual iage quality etrics are used to jointly optiize the halftone iage. This ethod is shown to be ore robust against any coon iage processing operations such as JPEG copression and histogra equalization when copared to prior ethods which do not tae into account the printing process.

3. OUR APPROACH At Purdue University we are developing forensic signatures for consuer printers based on nowledge of the printer process and the analysis of printed docuents resulting fro a given printer. Our intention is to develop these ethods for both injet (IJ) and electrophotographic (EP or laser) printers. We believe that secure printing is based on the concept that the printer output, the docuent, is an effective eans for identifying features of the printer. These features, which are printer specific, can be used for docuent security. For exaple, in the case of a suspected forgery we ideally should be able to tell what type of printer was used to create the docuent. In our wor at Purdue the focus is on extracting features fro the docuent and developing ethods to securely hide inforation in docuents based on these features. We propose to develop, 3 two strategies for printer identification: intrinsic and extrinsic signatures. We are developing iage analysis techniques to extract features fro a printed docuent that can be used as an intrinsic signature. The intrinsic signature will be detected by scanning the printed page with a high resolution scanner and treating it as a high resolution iage. We will then extract features fro the iage and use the to classify the docuent with respect to the printer. We will then use the features that describe the intrinsic signature to design the extrinsic signature. In contrast to currently used docuent wateraring ethods which ebed the waterar into the docuent before it is printed, we propose to create the waterar by odifying the print process, this is the extrinsic signature. This will be accoplished by odulating the process paraeters in the printer echanis to encode identifying inforation such as the printer serial nuber and date of printing. We believe that oving this ebedding step into the printer hardware will deter possible attepts to odify the hidden inforation before it is printed. If these steps were instead handled by a print driver or soe other software beforehand, then a alicious user could reverse engineer the ebedding software and re-engineer it to suit his own needs. 4. BANDING IN EP PRINTING Our current wor has focused on intrinsic feature detection of EP (laser) printers. In order to gain insight into the types of features that can be used to describe these printers, an understanding of the EP printing process is necessary. The first thing to note is that in the printed output fro any printer there exist defects caused by electroechanical fluctuations or iperfections in the print echanis. 3 Because these print quality defects are directly related to the printer echanis, they can also be viewed as an intrinsic signature or feature of the printer. 4.. Banding as an Intrinsic Feature Figure shows a side view of a typical EP printer. The print process has six steps. The first step is to uniforly charge the optical photoconductor (OPC) dru. Next a laser scans the dru and discharges specific locations on the dru. The discharged locations on the dru attract toner particles which are then attracted to the paper which has an opposite charge. Next the paper with the toner particles on it passes through a fuser and pressure roller which elt and peranently affix the toner to the paper. Finally a blade or brush cleans any excess toner fro the OPC dru. In EP printing, the ajor artifact in the printed output is banding which is defined as the artifacts that are due to quasiperiodic fluctuations in process direction paraeters in the printer. These are priarily due to fluctuations in the angular velocity of the OPC dru and result in non-unifor scan line spacing. This causes a corresponding fluctuation in developed toner on the printed page. 28 The appearance of the banding artifact is alternating light and dar bands perpendicular to the process direction (the direction the paper passes through the printer). The ain cause of banding is electroechanical fluctuations in the printer echanis, ostly fro gear baclash. Because these fluctuations are related to the gearing, the banding frequencies present in the printed page should directly reflect echanical properties of the printer. To easure the banding frequencies of an EP printer, idtone graylevel patches created with a line fill pattern were printed and analyzed. 3 Figure 2 shows a 25% line fill pattern. This pattern was printed on a set of EP printers and then each patttern was scanned at 24dpi. Each scanned iage, ig(i, j), was then projected

Cleaning Blade Charge Roller /6 inch 3/6 inch A Laser Bea F Fused Toner Particles Fuser Roller E OPC Dru D B C Developer Roller Paper Pressure Roller Process Direction Transfer Roller Figure. Diagra of the EP process: (A) charging, (B) exposure, (C) developent, (D) transfer, (E) fusing, (F) cleaning Figure 2. Midtone pattern 25% fill used for banding characterization horizontally to produce proj(i) = j ig(i, j) shown in Figure 3a. A Fourier analysis of the projection was then obtained and shown in Figure 3b which shows spies at 32 cycles/inch and 5 cycles/inch. Table shows a list of printers and their principle banding frequencies as found by this ethod. 6.3 x 5 Horizontal Projection of 25% Fill Pattern fro HP lj6mp 8 x 5 FFT of Projection of 25% Fill Patch for HP lj6mp 6.25 6 6.2 4 proj(i) 6.5 6. 6.5 6 Aplitude (graylevels) 2 8 6 5.95 4 5.9 2 5.85 2 4 6 8 2 4 i 5 5 2 25 Frequency (cycles/inch) (a) Projection of 25% fill patern fro HP lj6mp (b) FFT of the projection showing peas at 32 and 5 cycles/inch Figure 3. Projection and FFT analysis of 25% fill pattern printed on HP lj6mp Detection and easureent of the banding signal in docuents with large idtone regions, such as those with graphic art, can easily be done using ethods siilar to that used to produce Table. Our interest is to be able to detect these signals in text. This will require a different processing technique due to the low SNR of the banding signal with respect to the text and because of the liited nuber of cycles of the banding signal which can be captured in the length of one text character. 4.2. Banding Modelling and Analysis In this section we will describe our odel of banding and our initial analysis approach. Our ultiate goal is to reliably detect the banding frequencies fro a standard printed page. We believe it is iportant in soe of our forensic applications to be able to do this without using specific test docuents. 4.2.. Banding Modelling Our odel is shown in Figure 4. Two paths are shown in this figure, one for processing actual docuents, and another for processing ideal docuents. These are referred to as the actual path and the ideal path respectively.

Printer Model Principal Banding Frequencies (cycles/inch) HP LaserJet 5MP 37, 74 HP LaserJet 6MP 32, 5 HP LaserJet 27, 69 HP LaserJet 2 69 HP LaserJet 45 5, Sasung ML-45 6, 32,, 6 Table. Banding Frequencies for Several Printers. y actual (i,j) n^actual (i,j) Print Scan Morphological Filter Run Extractor Run Analysis PS File b ideal (i) PS > TIFF Add Ideal Banding Morphological Filter Run Extractor Run Analysis n ideal (i,j) y ideal (i,j) n^ideal (i,j) Figure 4. Banding Analysis Model A rando docuent generator is used to generate full page text docuents that we will use to characterize the printer. We used a odified version of a Marov chain rando text generator 29 (the Money ) which we trained on text available fro the Gutenberg project. 3 We used The Adventures of Sherloc Holes by Sir Arthur Conan Doyle. The progra first estiates the conditional probabilities of all the characters in the training data fro order to order 4. It then generates a user specified length of text following these transition probabilities and produces output into a L A TEX foratted file to allow easy foratting and conversion into PostScript (PS). We refer to this syste as the Forensic Money Text Generator (FMTG). Following the actual path, we start with a PS representation of a docuent generated using the FMTG described above. The PS file is first sent to a printer which will create the printed docuent. This docuent is then scanned at 24dpi and stored as a TIFF iage which is then processed to estiate the banding frequencies. The ideal path is used to construct a siple odel of the banding process and gain soe insight into how the banding anifests itself in the printed docuent. The PS representation of the docuent is first converted using ghostview into a TIFF iage at 24dpi, the sae resolution as is used in the scan step on the actual path. This iage can be thought of as the ideal text, n ideal (i, j), since it is binary with the bacground having graylevel, or white, and the text having graylevel 255, or blac. It is assued that =white and 255=blac. Next an ideal banding signal is added to the ideal text to generate the iage y ideal (i, j) which will be the ideal representation of the output of the scan process in the actual path. All processing and analysis is perfored on 2x2 pixel blocs. Figure 5 shows a bloc of ideal text. Let { n ideal 253, n (i, j) = ideal (i, j) belongs to a text character (), else where text characters have a graylevel of 253 and the bacground has graylevel. The reason for choosing not to represent the graylevel of the text as 255 will becoe apparent when we add the ideal banding signal. The origin is defined as the top-left corner of the bloc with the i axis going downward and the j axis going toward the right.

j ideal Current Slice n (i,596) 3 Slice n ideal (i,596) of Ideal Text n ideal (i,j) Inside of letter o 25 Graylevel 253 i Graylevel (=white 255=blac) 2 5 Top of letter o Botto of letter o 5 Top of letter l Botto of letter l n ideal (i,j), 2x2 pixel Bloc of Ideal Iage 2 4 6 8 2 i Figure 5. 2x2 pixel bloc of ideal text Figure 6. Slice fro a bloc of ideal text Each bloc is then broen into slices n ideal (i) for fixed j, an exaple of which is shown in Figure 6 corresponding to the 596th slice ared in Figure 5. Within each slice we define runs, r ideal (), which are sections of the slice that are part of a text character and long enough to capture at least one cycle of the sallest banding frequency to be estiated. The subscript denotes the run nuber, starting with = for the first run in slice, and ending with the last run in slice 2. The index denotes the position in the run starting fro =. Next we add an ideal banding signal to n ideal (i, j) to create y ideal (i, j), the ideal representation of a bloc of text in the actual path. We will assue that the printer in the ideal path is an HP lj5mp, and that only the 37 cycles/inch banding signal is present. Figure 7 shows an ideal banding signal with an aplitude of 2 graylevels and frequency 37 cycles per inch as defined by b ideal (i) = a sin(2πf i ), (2) dpi where a is the aplitude in graylevels, f is the frequency in cycles/inch, and dpi is the resolution of the signal in dots per inch. This signal only varies in the i direction because, as stated earlier, the banding signal is a print defect in the process direction only. The addition of the banding signal to the ideal bloc of text is a special process. The signal only anifests itself in printed areas eaning that in the case of the ideal text bloc it will only appear on top of a text character and not in the white bacground areas. To for the banded iage, y ideal (i, j), the following operation is perfored, { y ideal n (i, j) = ideal (i, j) + b(i), if n ideal (i, j) = 253 n ideal (3) (i, j), else where the banding signal is added only in printed areas. Figure 8 shows the sae slice seen in Figure 6 but now with the added banding signal. Figure 9a shows the banded run fro this slice. The average value of the run is approxiately 253, and the banding signal itself is seen as a graylevel variation between 25 and 255. 4.2.2. Banding Analysis Since it is the banding signal that is of interest, a easure of the SNR between the banding signal and the noise, which in this case is the text, ight give soe insight into what estiation techniques should be used. To siplify the situation, we will only exaine a single slice. Additionally we will assue that the text character profile is a square pulse with width 5 pixels and that the spacing between lines, or character profiles in the slice, is 3 pixels. This is representative of 2point text. Since the banding signal only appears in printed areas, we can safely exaine a single 5 pixel run containing the profile of one text charater. The SNR is the ratio of

3 Ideal Banding Signal b(i)=a*sin(2*π*i*f/dpi), f=37(cycles/inch), dpi=24(dots/inch) 3 ideal ideal ideal Slice y (i,596) = n (i,596) b (i) Graylevel 255 Inside of letter o 2 25 Graylevel 25 Aplitude (in Graylevels) Graylevel (=white 255=blac) 2 5 Top of letter o One Run r ideal () length 222 Botto of letter o 2 5 Top of letter l Botto of letter l 3 2 4 6 8 2 i 2 4 6 8 2 i Figure 7. Ideal banding signal with aplitude 2 and frequency 37 cycles/inch Figure 8. Slice fro a bloc of ideal text with banding signal the energy in the banding signal divided by the energy in the text profile. SNR = E banding E textprofile = 5 i= b2 (i) 5 = 6.2285 8 i= 2532 Clearly trying to extract the signal directly fro the slices will be probleatic, hence we process the runs instead of the slices. Extracting the runs fro the ideal text is trivial since the text and bacground graylevel values are well defined. The siplest ethod in this case is to use a thresholding technique to loo for runs of pixels whose values are greater than. For actual text this will not wor since the edges of the text are not well defined and the bacground is not unifor graylevel. To reedy this proble, we first process the bloc of text with a 5x5 close-open orphological filter. This operation will produce a as which ars the locations of the text characters with graylevel 255, and the bacground with graylevel. This as is then used to extract the runs and their locations are then apped bac onto the original bloc for analysis. Starting with the extracted run fro an ideal slice, such as that shown in Figure 9a, first the ean of the run is reoved foring () = r ideal () E[r ideal ()]. (5) r ideal The ean shift reoves the DC coponent of the signal, as shown in Figure 9b. Next the signal is windowed using a raised cosine window foring r ideal w () = r ideal () cos(π (4) N π 2 ) (6) where N is the length of the run r ideal. Again the ean is subtracted fro the signal to reove its DC coponent to for r ideal w () = r ideal w () E[r ideal w ()]. (7) This process of preparing each run for frequency analysis is shown in Figure 9 along with the FFT of r ideal w (). In this ideal case a pea is clearly seen centered at 36.92 cycles/inch, very near to the actual ebedded banding frequency of 37 cycles/inch. This sae ethodology is used to analyze actual text in our initial experients. Figure shows a bloc of actual text as well as a closeup of the letter l fro the first line of text in the bloc. Notice in these iages that the actual text is very noisy and the characters and bacground are not unifor in graylevel. Shown in Figure are three runs fro this text bloc. The sharp pea in the first run, and the sharp dip in the second run correspond to an excess and deficiency of developed toner on the docuent. The first run has a pea near the expected banding frequency of 37 cycles/inch although the axiu aplitude was easured at 2.56

5 5 2 25 3 Plot of Run r ideal () fro y ideal (i,596), ean=253.387 3 Plot of Run r ideal, () 3 Plot of Windowed Run r ideal, w () and Raised Cosine Window length 222, ean=.262 25 2 2 Aplitude (in Graylevels) 2 5 Aplitude (in Graylevels) Aplitude (in Graylevels) 5 2 2 3 5 5 2 25 3 5 5 2 25 (a) (b) (c) 3 ideal w Plot of Windowed Run r.4 ideal w One sided Fourier Transfor of r (), axiu at 36.92 cycles/inch 2.2 Aplitude (in Graylevels) Aplitude.8.6.4 2.2 3 5 5 2 25 2 3 4 5 6 7 8 9 Frequency (cycles/inch) (d) (e) Figure 9. Processing of runs: a) Original Run r ideal (), b) r ideal () = r ideal cos(π π ), d) N 2 rideal w () = r ideal w () E[r ideal w ], e) FFT of r ideal w () () E[r ideal ], c) r ideal w () = r ideal () (a) (b) Figure. a) Bloc of text fro HP lj5mp b) Letter l fro first line cycles/inch. The second run has a pea located at 28.3 cycles/inch. The third run has a pea at approxiately 42 cycles/inch. Fro our analysis of the runs in actual text, we believe that it is possible to estiate the banding frequency using our odel. We are currently exaining wea signal detection ethods for extracting the banding frequency.

Graylevel 9 8 7 6 5 4 r 5 (), location i=58 j=898, length 2, ean=.848e+2 3 5 5 2 25 Graylevel r 5 (), location i=88 j=473, length 2, ean=.876857e+2 2 98 96 94 92 9 88 86 84 82 8 5 5 2 25 Graylevel 9 88 86 84 82 8 78 r 2 (), location i=85 j=478, length 22, ean=.84366e+2 76 5 5 2 25 (a) (b) (c) Aplitude 3 2.5 2.5 R 'w' 5 (f), location i=58 j=898, length 2 ax 2.56.5 37 cpi 2 3 4 5 6 7 8 9 Frequency (cycles/inch) Aplitude.6.4.2.8.6.4.2 R 'w' 5 (f), location i=88 j=473, length 2 ax 28.3 37 cpi 2 3 4 5 6 7 8 9 Frequency (cycles/inch) Aplitude.8.7.6.5.4.3.2. R 'w' 2 (f), location i=85 j=478, length 22 ax 8.2 37 cpi 2 3 4 5 6 7 8 9 Frequency (cycles/inch) (d) (e) (f) Figure. a-c) Runs fro actual text Figure a, d-f) Corresponding FFT of the processed runs a-c 4.3. Banding as an Extrinsic Signature We have also exained the possibility of using banding as an extrinsic feature. Prior wor has been reported on reducing the banding artifacts in the print process by controlling either the OPC dru velocity itself or 28, 3 odulating the laser pulse width as a function of the OPC dru velocity. In the first ethod, a closed loop controller varies the velocity of the otor controlling the OPC dru based on the real tie velocity that is easured using an encoder. The second ethod uses the fact that the velocity of the OPC dru dictates the scan line spacing assuing that the laser scanning speed is constant. When the OPC dru velocity is low, the scan lines are closer together and so the pulse width is decreased to reduce the aount of developed toner on the printed docuent. When the OPC dru velocity is high, the scan lines will be further apart and so the pulse width is increased to increase the aount of developed toner on the printed docuent. Both ethods have been shown to wor very well. Using either of the banding reduction ethods, it would be possible to odulate any existing banding frequencies in the print process to encode inforation. Additionally we can inject arbitrary artificial banding frequencies into a docuent. We have been successful in ebedding artificial banding frequencies into a docuent by using the first ethod. Because of the electroechanical liitations iposed by the otor controlling the OPC, we have not been able to inject frequencies higher than 5 cycles/inch. We are currently woring with the second ethod which should allow the injection of frequencies above cycles/inch. 5. CONCLUSION Using print quality defects as an intrinsic signature of a printer is possible as shown for banding artifacts in EP printers. We believe that by successfully extracting the banding features fro a docuent it is possible to classify the docuent based on the specific ae and odel of device which created it. We need to better odel this process by eploying ore sophisticated statistical analysis tools.

REFERENCES. E. J. Delp, Is your docuent safe: An overview of docuent and print security, Proceedings of the IS&T International Conference on Non-Ipact Printing, San Diego, California, Septeber 22. 2. A. M. Esicioglu and E. J. Delp, An overview of ultiedia content protection in consuer electronics devices, Signal Processing : Iage Counication, vol. 6, pp. 68 699, 2. 3. A. M. Esicioglu, J. Town, and E. J. Delp, Security of digital entertainent content fro creation to consuption, Signal Processing : Iage Counication, vol. 8, no. 4, pp. 237 262, April 23. 4. C. I. Podilchu and E. J. Delp, Digital wateraring: Algoriths and applications, IEEE Signal Processing Magazine, vol. 8, no. 4, pp. 33 46, July 2. 5. M. Barni, C. I. Podilchu, F. Bartolini, and E. J. Delp, Waterar ebedding: hiding a signal within a cover iage, IEEE Counications Magazine, vol. 39, no. 8, pp. 2 8, August 2. 6. B. Macq, J. Dittann, and E. J. Delp, Bencharing of iage wateraring algoriths for digital rights anageent, Proceedings of the IEEE, 24, to appear in. 7. R. W. Wolfgang, C. I. Podilchu, and E. J. Delp, Perceptual waterars for digital iages and video, Proceedings of the IEEE, vol. 87, no. 7, July 999, pp. 8 26. 8. I. J. Cox, M. L. Miller, and J. A. Bloo, Digital Wateraring. San Francisco, CA: Morgan Kaufann Publishers, 22. 9. R. L. van Renesse, Paper based docuent security - a review, IEE European Conference on Security and Detection, no. 437, pp. 75 8, April 997.. J. Picard, Digital authentication with copy detection patterns, Proceedings of the SPIE International Conference on Optical Security and Counterfeit Deterrence Techniques V, vol. 53, January 24.. R. Steenbli and M. Hurt, Unison (t) icrooptic security fil, Proceedings of the SPIE International Conference on Optical Security and Counterfeit Deterrence Techniques V, vol. 53, January 24. 2. R. L. Renesse, Optical Docuent Security. Boston: Artech House, 998. 3. G. N. Ali, P.-J. Chiang, A. K. Miilineni, J. P. Allebach, G. T. Chiu, and E. J. Delp, Intrinsic and extrinsic signatures for inforation hiding and secure printing with electrophotographic devices, Proceedings of the IS&T International Conference on Non-Ipact Printing, vol. 9, New Orleans, LA, Septeber 23, pp. 5 55. 4. J. T. Brassil, S. Low, and N. F. Maxechu, Copyright protection for the electronic distribution of text docuents, Proceedings of the IEEE, vol. 87, no. 7, July 999, pp. 8 96. 5. J. T. Brassil, S. Low, N. F. Maxechu, and L. O Goran, Electronic aring and identification techniques to discourage docuent copying, IEEE Journal on Selected Areas in Counication, vol. 3, no. 8, pp. 495 54, October 995. 6. S. H. Low and N. F. Maxechu, Capacity of text aring channel, IEEE Signal Processing Letters, vol. 7, no. 2, pp. 345 347, Deceber 2. 7. T. N. Pappas, J. P. Allebach, and D. L. Neuhoff, Model-based digital halftoning, IEEE Signal Processing Magazine, vol. 2, no. 4, pp. 4 27, July 23. 8. M. S. Fu and O. C. Au, Data hiding wateraring for halftone iages, IEEE Transactions on Iage Processing, vol., no. 4, pp. 477 484, April 22. 9. G. Shara and S. Wang, Show-through wateraring of duplex printed docuents, Proceedings of the SPIE International Conference on Security, Steganography, and Wateraring of Multiedia Contents VI, vol. 536, January 24. 2. Z. Fan, Inforation ebedding using two-layer conjugate screening, Proceedings of the SPIE International Conference on Optical Security and Counterfeit Deterrence Techniques V, vol. 53, January 24. 2. S. V. Voloshynovsiy, F. Deguillaue, and T. Pun, Multibit digital wateraring robust against local nonlinear geoetrical distortions, Proceedings of the IEEE International Conference on Iage Processing, pp. 999 2, October 2. 22., Content adaptive wateraring based on a stochastic ultiresolution iage odeling, Proceedings of the European Signal Processing Conference, Septeber 2.

23. S. V. Voloshynovsiy, F. Deguillaue, S. Pereira, and T. Pun, Optial adaptive diversity wateraring with channel state estiation, Proceedings of the SPIE International Conference on Security and Wateraring of Multiedia Contents III, vol. 434, San Jose, California, August 2, pp. 673 685. 24. F. Deguillaue, S. V. Voloshynovsiy, and T. Pun, Method for the estiation and recovering fro general affine transfors in digital wateraring applications, Proceedings of the SPIE International Conference on Security and Wateraring of Multiedia Contents IV, vol. 4675, April 22, pp. 33 322. 25. D. Kacer and J. P. Allebach, Joint halftoning and wateraring, IEEE Transactions on Signal Processing, vol. 5, no. 4, pp. 54 68, April 23. 26. D. J. Lieberan and J. P. Allebach, A dual interpretation for direct binary search and its iplications for tone reproduction and texture quality, IEEE Transactions on Iage Processing, vol. 9, no., pp. 95 963, Noveber 2. 27. I. J. Cox, J. Kilian, F. T. Leighton, and T. Shaoon, Secure spread spectru wateraring for ultiedia, IEEE Transactions on Iage Processing, vol. 6, no. 2, pp. 673 687, Deceber 997. 28. G.-Y. Lin, J. M. Grice, J. P. Allebach, G. T.-C. Chiu, W. Bradburn, and J. Weaver, Banding artifact reduction in electrophotographic printers by using pulse width odulation, Journal of Iaging Science and Technology, vol. 46, no. 4, pp. 326 337, July/August 22. 29. T. McConnell. (999) Money: a arov chain rando text generating progra. [Online]. Available: http://barnyard.syr.edu/ tc/ 3. [Online]. Available: http://www.gutenberg.net/ 3. C.-L. Chen and G. T.-C. Chiu, Banding reduction in electrophotographic process, Proceedings of the IEEE/ASME International Conference on Advanced Intelligent Mechatronics, vol., July 2, pp. 8 86.