Forensic Authentication of Data Bearing Halftones
Stephen Pollard
1
, Robert Ulichney
2
, Matthew Gaubatz
3
and
4
Steven Simske
1
Hewlett Packard Labs, Bristol, U.K.
2
Hewlett Packard Labs, Andover, Massachusetts, U.S.A.
3
Hewlett Packard Labs, Bellevue, Washington, U.S.A.
4
Hewlett Packard Labs, Fort Collins, Colorado, U.S.A.
Keywords: Forensic Printing, Stegatones, Image Registration, Gabor Filters, Biometrics.
Abstract: This paper introduces a practical system for combining overt, covert and forensic information in a single,
small printed feature. The overt “carrier” feature need not be a dedicated security mark such as a 2D or
color barcode, but can instead be integrated into a desirable object such as a logo as part of the aesthetically-
desired layout using steganographic halftones (Stegatones). High-resolution imaging in combination with
highly accurate and robust image registration is used to recover, simultaneously, a unique identity suitable
for associating a unique print with an on-line database and a unique forensic signature that is both tamper
and copy sensitive.
1 INTRODUCTION
Counterfeiting, warranty fraud, product tampering,
smuggling, product diversion and other forms of
organized deception are driving the need for
improved brand protection. The potential for
security printing and imaging to provide an
extremely cost-effective forensic level of
authentication is well-recognized (Pizzanelli, 2009).
There are also a number of instances in which
embedding data in hard copy is desired, but overt
marks such as bar codes would damage the
aesthetics of the document. The novel method,
outlined in this paper, simultaneously addresses both
of these needs by combining forensics and
steganographic halftoning (Ulichney et al., 2010) on
the same printed object, and describes a system for
both encoding and decoding such objects.
In order to perform a forensic authentication of
printed material, it is necessary use an image
resolution sufficient to expose unique properties of
the print that are extremely difficult, if not
impossible on a regular paper substrate, to reproduce
or copy (Pollard et al., 2010). For the majority of
printing technologies, these properties result
naturally from the stochastic nature of the print
process itself and its interaction with the underlying
structural properties of the substrate material on
which ink is printed. As such they represent a unique
fingerprint that can be used to authenticate
individually printed items such as labels, documents,
product packaging and monetary notes.
Previously (Pollard et al., 2012) a method
derived from iris recognition (Daugman, 1993) has
been used to derive a general area-based print
signature that can be applied to halftones images and
thus affords general utility and applicability for
forensic print authentication. Here that idea is
extended to show that the methodology developed
for regular halftones is applicable to steganographic
halftone, or Stegatone, images where the content of
the original halftone has been modulated, in a
manner unknown to the decoding system, to carry
extra covert information. Most importantly, the
image alignment strategy on which the method is
founded is not disrupted by the introduction of
unknown deformations in the printed material.
Furthermore, despite the small extent of the
stegatones used in our experiments (4mm on a side),
they are able to encode sufficient bit data to be a
practical alternative overt 2D barcode alternatives
such as Data Matrix or QR-Codes.
2 METHOD
The Stegatone encoding system outlined in Figure 1
allows the creation of a secure hardcopy document
with an embedded payload along with the filing of
its forensic signature in a registry located on a
109
Pollard S., Ulichney R., Gaubatz M. and Simske S..
Forensic Authentication of Data Bearing Halftones.
DOI: 10.5220/0004296201090113
In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2013), pages 109-113
ISBN: 978-989-8565-48-8
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
server.
T
p
rinted
(
hand co
r
is modi
f
able to
u
directly
identify
t
Fig
2.1
S
A Stega
t
input i
m
that tra
n
p
aper,
w
as the
m
represen
or natur
a
a natura
l
92x92 p
i
mule i
m
size in F
Refe
r
standar
d
mule i
m
referenc
e
data car
r
0-
b
it ca
r
Cells” b
e
to aid
a
represen
can shif
t
referenc
e
T
he selected
(
e.g. automat
i
r
ner of every
f
ied to inclu
d
u
niquely iden
t
or referenti
a
t
he user, time
g
ure 1: Forensi
c
S
te
g
atone
G
t
one generat
o
m
age called a
“
n
sports the
p
w
e are using
a
m
ule to carry
n
t any type of
a
l images (th
o
l
image is so
m
i
xel (4mm sq
u
m
ages are sho
w
igure 2.
r
ence halfto
n
d
clustered-do
t
m
age. All hal
f
e
map as eit
h
r
iers. These c
e
r
riers, colore
d
ecause they
a
a
lignment. T
h
n
t 1, 2 & 3-
bi
t
in one of 2,
4
e
cells beca
u
targe
t
(or m
u
i
cally inserte
d
page during
d
e payload i
n
t
ify the docu
m
a
lly (throug
h
and location
c
Stegatone enc
o
G
eneration
o
r takes a dat
“
mule” becau
s
p
ayload whe
n
a
4mm squar
e
the payload
.
object includ
i
o
ugh at this s
c
m
ewhat limit
e
u
are at 600dp
i
w
n at many
t
n
es (e.g.
F
t
halftones g
e
f
tone cells a
r
h
er 0-bit, 1-
b
e
lls are depic
t
d
black, are
c
a
re unchange
d
h
e red, gree
n
i
t carriers res
p
4
and 8 positi
o
u
se they are
u
le) image
w
d
in the top
r
t
he print pro
c
n
formation th
a
m
ent and – e
i
h
the registr
y
of the print.
o
ding system.
a
payload an
d
s
e it is the ve
h
n
printed. In
e
grayscale i
m
.
This image
i
ng, glyphs, l
o
c
ale the conte
n
e
d). Two exa
m
i
) continuous
t
imes their a
c
F
igure 3(a))
e
nerated fro
m
r
e classified
b
it, 2-
b
it, or
3
t
ed in Figure
3
c
alled “Refer
e
d
and can be
u
n
and blue
c
p
ectively (as
o
ns). Cells c
a
too large t
o
w
hen
r
ight
c
ess)
a
t is
i
ther
y
) –
d
an
h
icle
this
m
age
can
o
gos
n
t of
m
ple
tone
c
tual
are
m
the
in a
3
-bit
3
(b).
e
nce
u
sed
c
ells
they
a
n be
o
be
shi
f
int
e
if t
h
Th
e
shi
f
ca
p
res
p
Fig
sh
o
2.
2
Th
e
b
y
an
d
(D
a
has
co
m
as
i
b
ill
aut
h
ou
r
im
a
Im
a
hig
h
are
f
ted or too
e
restingly, we
h
eir unaltere
d
e
payload is
f
ts of the h
a
p
acity of thes
p
ectively, fo
r
u
re 2. The pa
y
o
wn in 3(c).
2
Forensi
e
approach fo
Daugman in
d
expanded
a
ugman, 200
6
become the
m
mercial bio
m
i
t does, the a
b
ions of iris p
a
Iris recogni
t
h
entication t
a
images are
c
a
ging devic
e
a
ging Device
;
h
resolution (
a
captured usi
n
small to be
can force ce
l
d
shape is hel
p
encoded by
m
a
lftone cluste
r
e examples
a
r
the left a
n
y
load is enco
d
Figure 2
.
Figure 3
.
c
Print Si
g
l
lows a meth
o
his 1993 pa
p
on in su
b
6
); (Daugman
,
b
ackbone of
m
m
etric recogn
i
b
ility to robu
s
a
tterns.
t
ion differs
sk in three i
m
c
aptured usin
g
DrCID (
D
;
Adams, 20
1
a
bout 7900dp
n
g traditional
o
detected;
b
l
ls to be refer
e
pful during
a
means of si
n
rs. The data
a
re 447 and
n
d right exa
m
d
ed in the St
e
.
.
g
nature
o
dology first
p
er on iris r
e
b
sequent pu
b
,
2007). This
many gover
n
i
tion systems;
s
tly discrimi
n
from the
S
m
portant rega
r
g
a specializ
e
D
yson Rela
y
1
0) at an al
m
p
i), whereas i
r
optics and th
u
b
ut, more
e
nce cells
lignment.
n
gle pixel
carrying
349 bits,
m
ples in
gatone as
proposed
cognition
b
lications
approach
ment and
offering,
n
ate many
S
tegatone
r
ds. First,
e
d contact
y
CMOS
m
ost fixed
r
is images
u
s vary in
VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications
110
size over a small but significant range. Second, parts
of the iris are not properly imaged due to either
obscuration (by the eyelids or the eye-lashes) or
specular reflections of the near-infrared light
sources. Thus encoded features extracted from these
regions must be robustly and accurately excluded
from the statistical comparison process. Print
images, on the other hand, do not generally suffer
such imperfections and the whole of the encoded
sequence can be used. Finally, unique iris features
can be encoded across a wide range of spatial
frequencies while the random perturbations
associated with printed halftones are more limited.
For both Stegatone decoding and print signature
extraction it is important to accurately and with good
repeatability be able to register the captured
Stegatone image as shown in Figures 3(d). In this
work Stegatone patterns are registered using multi-
scale gradient descent (Bouguet, 1999) derived from
the well-established Lucas and Kanade (1981)
method. For the multi-scale representation we
normalize band pass filters (difference of successive
Gaussian filtered images) to have unit standard
deviation in order to minimize the difference
between the stylized scaled (13x from 600 to
7900dpi) half-tone images and the printed and
captured Stegatones that are closely related to them.
Following Daugman’s methodology, the random
signal is demodulated to extract its phase
information using quadrature 2-D Gabor wavelets.
In our case the Gabor filters use Cartesian
coordinates and not the polar coordinates used for
iris biometrics. That is
ℎ
,
=
,
,
.
where
=
⁄
⁄
,
=
ignoring orientation: where h
{Re, Im}
is a complex
valued bit whose real and imaginary parts are either
1 or 0 depending on the sign of the 2-D integral; I(x,
y) is the warped raw image; α and β are size
parameters of the Gaussian envelope; the parameter
ω
0
is the spatial frequency of the filter. There is an
additional orientation parameter
θ
0
which is ignored
in this formulation for simplicity. Thus, for all
samples each wavelet provides two bits towards the
phase encoding that describes the random elements
of the printed halftone. Samples can be combined
spatially over an M x M grid and through the choice
of filter control parameters – notably frequency
ω
and orientation
θ
.
3 RESULTS & CONCLUSIONS
We have printed a number (>8) of identical halftone
and Stegatone images on 3 identical HP4345 Laser
Printers. There are two versions of an HP logo
(labeled Logo1 and Logo2 with white and grey
backgrounds respectively) and the Rainbow Bridge.
Each print is captured twice (using different imaging
devices) in order to compare the fractional Hamming
Distance (HD) scores of valid matches with those of
the binomially distributed statistically independent
false matches.
First let’s compare the statistical properties of
Stegatone derived Gabor signature profiles with
those derived from halftones as previously reported
in Pollard et al (2012). The crucial difference is that
the same digital halftone model is used to register
both the halftone print (which was derived from it
directly) and the Stegatone which includes small but
significant deviations from the original halftone;
which are unknown at the start of the decoding
process. In Table 1 fractional Hamming distance
statistics are compared for each of the three printed
images. Each row represents false comparisons
amongst all collected halftone (HT) and Stegatone
(ST) images. For this test, there were 24 such images
(276 comparisons) for all cases except the original
Rainbow halftone images reported in the earlier
paper for which there were 48 images (1128
comparisons). In every case, a single Gabor filter
was used with
λ
= 8 pixels and two sampling
densities M = 32 (which leads to a 2K bits/256 bytes
code used for iris biometrics) and M = 80 (beyond
which recognition rates were found to plateau). In all
cases, except M = 80 for the second HP logo, the
mean and standard deviations of Stegatone images
compared to their halftone equivalents were
sufficiently similar as to be considered the same
within the 95% confidence limit of the t-Test.
Table 1: Hamming Distance Statistics.
ForensicAuthenticationofDataBearingHalftones
111
The results in Table 1 show that the false match
distribution statistics are not significantly altered by
the change from halftone to Stegatone printing. This
is not very surprising as the frequency content of
each image type is significantly the same.
Furthermore, the collection of false match statistics
is not dependent on high accuracy registration and
so is not likely to be affected by mismatch between
the digital halftone used to register the Stegatone
image. Of more interest is the effect this mismatch
has on the Hamming distance of correct matches.
Figure 4 shows a scatter plot of fractional Hamming
distance for 20 correctly matching image pairs for
the Rainbow Bridge Stegatone (
λ
= 8; M = 80) along
with the (rotated) probability density function (PDF)
for false matches. As can be seen clearly from this
figure the probability of any of the correct matches
being generated by chance is very low indeed. In
fact the average z-score of false positives for the
Rainbow Bridge is 57.95 which corresponds to a
markedly small probability of 4x10
-732
.
Figure 4.
Using the average z-score as a representative
shorthand for the statistical robustness of the
forensic print signature, Figure 5 compares halftone
and Stegatone values for the conditions presented in
Table 1. Robustness is clearly maintained for all
conditions and image types. Note that the longer
phase code (M = 80) results in much greater
statistical robustness. This improvement can be
increased further by combining extra Gabor filter
frequencies and orientations.
Thus it is possible to use a single small 4mm
square Stegatone print and capture with a high
resolution imaging device to provide both covert
data encoding (raw error rates for the best printer
were 10%, 6% and 1% for the respectively for the
Logo1, Logo2 and Rainbow stegatones) and a
unique forensic print signature that exploits the
stochastic nature of the print process and underlying
surface substrate of the paper on which it is printed.
Figure 5: Mean z-scores for valid matches (
λ
= 8; M = [32,
80]) for both Stegatone (ST) and halftone (HT) image
data.
Despite their modest size, Stegatones of this kind are
able to encode considerable amounts of data; in fact
the Rainbow Bridge example is able to robustly
encode 256 bits which is at least comparable to the
highest resolution 2D barcode of this size. In fact 2D
barcodes rarely, if ever, encode more than 150 bytes
per square cm, meaning a 4 x 4 mm barcode would
be no more than 200 bits, easily outdistanced by the
examples herein.
Using the Gabor phase coding approach
halftones and Stegatones of this size are able to
practically discriminate an almost infinite number of
printed instances. While iris biometrics limited the
code size to 2K bits, the high resolution (7200ppi)
images used in these experiments allows us to
greatly extend the code length (real and effective)
through higher sampling frequency. In fact it is
possible to increase this yet further by adding more
independent Gabor components at other frequencies
and orientations to achieve exceptional coding
efficiency (albeit at greater memory requirement for
the stored data).
REFERENCES
Adams, G., 2010, Handheld Dyson Relay Lens for Anti-
Counterfeiting, IEEE IST.
Bouguet, J-Y., 1999, Pyramid Implementation of Lucas
Kanade Feature Tracker: Description of the algorithm,
OpenCV Documents, Intel Corporation,
Microprocessor Research Lab.
Daugman, J. G., 1993, High confidence visual recognition
of persons by a test of visual phase information, IEEE
PAMI, 15(11).
Daugman, J., 2006, Probing the uniqueness and
randomness of IrisCodes: results from 200 billion
comparisons, Proc. IEEE, 94(11).
Daugman, J., 2007, New methods in iris recognition, IEEE
SMC, 37(5).
VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications
112
Lucas, B., Kanade, T., 1981, An iterative image
registration technique with an application to stereo
vision, IJCAI.
Pollard, S., Simske, S., Adams G., 2010, Model based
print signature profile extraction for forensic analysis
of individual text glyphs, IEEE WIFS.
Pollard, S., Simske, S., Adams G, 2012, Print Biometrics:
Recovering Forensic Signatures from Halftone
Images, IAPR ICPR.
Pizzanelli, D., 2009, The Future of Anti-Counterfeiting,
Brand Protection and Security Packaging V, Pira
International, Leatherhead, UK.
Ulichney, R., Gaubatz, M., Simske, S., 2010 Encoding
information in clustered-dot halftones, IS&T NIP26.
ForensicAuthenticationofDataBearingHalftones
113
