Usenix04

On User Choice in Graphical Password Schemes
Darren Davis Fabian Monrose Michael K. Reiter
Johns Hopkins University Carnegie Mellon University
{ddavis,fabian}@cs.jhu.edu reiter@cmu.edu

Abstract would be able to remember stronger passwords of a
graphical nature.

Graphical password schemes have been proposed as In this paper we study a particular facet of graphical
an alternative to text passwords in applications that password schemes, namely the strength of graphi-
support graphics and mouse or stylus entry. In cal passwords chosen by users. We note that not
this paper we detail what is, to our knowledge, the all graphical password schemes prescribe user cho-
largest published empirical evaluation of the effects sen passwords (e.g., [24]), though most do (e.g., [2,
of user choice on the security of graphical password 8, 3, 4, 7]). However, all of these schemes can be
schemes. We show that permitting user selection of implemented using either system-chosen or user-
passwords in two graphical password schemes, one chosen passwords, just as text passwords can be
based directly on an existing commercial product, user-chosen or system-chosen. As with text pass-
can yield passwords with entropy far below the the- words, there is potentially a tradeoff in graphical
oretical optimum and, in some cases, that are highly passwords between security, which benefits by the
correlated with the race or gender of the user. For system choosing the passwords, and usability and
one scheme, this effect is so dramatic so as to ren- memorability, which benefit by permitting the user
der the scheme insecure. A conclusion of our work to choose the password.
is that graphical password schemes of the type we
study may generally require a different posture to- Our evaluation here focuses on one end of this
ward password selection than text passwords, where spectrum, namely user chosen graphical passwords.
selection by the user remains the norm today. The graphical password schemes we evaluate are
a scheme we call “Face” that is intentionally very
closely modeled after the commercial PassfacesTM
scheme [3, 24] and one of our own invention (to our
1 Introduction knowledge) that we call the “Story” scheme. In the
Face scheme, the password is a collection of k faces,
each chosen from a distinct set of n > 1 faces, yield-
The ubiquity of graphical interfaces for applications, ing nk possible choices. In the Story scheme, a pass-
and input devices such as the mouse, stylus and word is a sequence of k images selected by the user
touch-screen that permit other than typed input, to make a “story”, from a single set of n > k im-
has enabled the emergence of graphical user authen- ages each drawn from a distinct category of image
tication techniques (e.g., [2, 8, 4, 24, 7, 30]). Graphi- types (cars, landscapes, etc.); this yields n!/(n − k)!
cal authentication techniques are particularly useful choices. Obviously, the password spaces yielded by
when such devices do not permit typewritten input. these schemes is exhaustively searchable by a com-
In addition, they offer the possibility of providing a puter for reasonable values of k and n (we use k = 4
form of authentication that is strictly stronger than and n = 9), and so it relies on the authentication
text passwords. History has shown that the dis- server refusing to permit authentication to proceed
tribution of text passwords chosen by human users after sufficiently many incorrect authentication at-
has entropy far lower than possible [22, 5, 9, 32], tempts on an account. Nevertheless, an argument
and this has remained a significant weakness of user given to justify the presumed security of graphical
authentication for over thirty years. Given the fact passwords over text passwords in such environments
that pictures are generally more easily remembered is the lack of a predefined “dictionary” of “likely”
than words [23, 14], it is conceivable that humans choices, as an English dictionary provides for En-

glish text passwords, for example (c.f., [8, Section The rest of this paper is structured as follows. We
3.3.3]). describe related work in Section 2. In Section 3
we describe in more detail the graphical password
For our study we utilize a dataset we collected dur- schemes that we evaluate, and discuss our data
ing the fall semester of 2003, of graphical password sources and experimental setup. In Section 4 we in-
usage by three separate computer engineering and troduce our chosen security measures, and present
computer science classes at two different universi- our results for them. In Section 5 we discuss issues
ties, yielding a total of 154 subjects. Students used and findings pertinent to the memorability of the
graphical passwords (from one of the two schemes two schemes. Finally, we conclude in Section 6.
above) to access their grades, homework, homework
solutions, course reading materials, etc., in a man-
ner that we describe in Section 3.2. At the end
of the semester, we asked students to complete an
exit survey in which they described why they picked
2 Related Work
the faces they did (for Face) or their chosen sto-
ries (for Story) and some demographic information
about themselves. This work, and in particular our investigation of the
Face scheme, was motivated in part by scientific lit-
Using this dataset, in this paper we evaluate the erature in psychology and perception. Two results
Face and Story schemes to estimate the ability of documented in the psychological literature that mo-
an attacker to guess user-chosen passwords, possibly tivated our study are:
given knowledge of demographic information about
the user. As we will show, our analysis suggests
that the faces chosen by users in the Face scheme • Studies show that people tend to agree about
is highly affected by the race of the user, and that the attractiveness of both adults and children,
the gender and attractiveness of the faces also bias even across cultures. (Interested readers are
password choice. As to the latter, both male and referred to [10] for a comprehensive literature
female users select female faces far more often than review on attractiveness.) In other words, the
male faces, and then select attractive ones more of- adage that “beauty is in the eye of the be-
ten than not. In the case of male users, we found holder,” which suggests that each individual
this bias so severe that we do not believe it possible has a different notion of what is attractive, is
to make this scheme secure against an online attack largely false. For graphical password schemes
by merely limiting the number of incorrect password like Face, this raises the question of what in-
guesses permitted. We also quantify the security of fluence general perceptions of beauty (e.g, fa-
the passwords chosen in the Story scheme, which cial symmetry, youthfulness, averageness) [1, 6]
still demonstrates bias though less so, and make rec- might have on an individual’s graphical pass-
ommendations as to the number of incorrect password choices. In particular, given these a pri-
word attempts that can be permitted in this scheme ori perceptions, are users more inclined to chose
before it becomes insecure. Finally, we benchmark the most attractive images when constructing
the memorability of Story passwords against those their passwords?
of the Face scheme, and identify a factor of the Story
scheme that most likely contributes to its relative • Studies show that individuals are better able to
security but also impinges on its memorability. recognize faces of people from their own race
than faces of people from other races [31, 20,
On the whole, we believe that this study brings into 11, 29]. The most straightforward account of
question the argument that user-chosen graphical the own-race effect is that people tend to have
passwords of the type we consider here are likely to more exposure to members of their own racial
offer additional security over text passwords, unless group relative to other-race contact [31]. As
users are somehow trained to choose better pass- such, they are better able to recognize intra-
words, as they must be with text passwords today. racial distinctive characteristics which leads to
Another alternative is to utilize only system-chosen better recall. This so-called “race-effect” [13,
passwords, though we might expect this would sacri- 15] raises the question of whether users would
fice some degree of memorability; we intend to eval- favor members of their own race when selecting
uate this end of the spectrum in future work. images to construct their passwords.

To the best of our knowledge, there has been no
prior study structured to quantify the influence of
the various factors that we evaluate here, including
those above, on user choice of graphical passwords,
particularly with respect to security. However, prior
reports on graphical passwords have suggested the
possibility of bias, or anecdotally noted apparent
bias, in the selection or recognition of passwords.
For example, a document [24] published by the cor-
poration that markets PassfacesTM makes reference
to the race-effect, though stops short of indicating
any effect it might have on password choice. In a
study of twenty users of a graphical password sys-
tem much like the Story scheme, except in which the
password is a set of images as opposed to a sequence,
several users reported that they did not select pho-
tographs of people because they did not feel they
could relate personally to the image [4]. The same
study also observed two instances in which users se-
lected photographs of people of the same race as Figure 1: In the Face scheme, a user’s password is a
themselves, leading to a conjecture that this could sequence of k faces, each chosen from a distinct set
play a role in password selection. of n > 1 faces like the one above. Here, n = 9, and
images are placed randomly in a 3 × 3 grid.
The Face scheme we consider here, and minor vari-
ants, have been the topic of several user studies fo-
black or white male or female, or an Asian, black or
cused on evaluating memorability (e.g., [34, 27, 28,
white male or female model. This categorization is
3]). These studies generally support the hypothe-
further discussed in Section 3.1. For our evaluation
sis that the Face scheme and variants thereof of-
we choose k = 4 and n = 9. So, while choosing her
fer better memorability than text passwords. For
password, the user is shown four successive 3 × 3
instance, in [3], the authors report results of a
grids containing randomly chosen images (see Fig-
three month trial investigation with 34 students that
ure 1, for example), and for each, she selects one im-
shows that fewer login errors were made when us-
age from that grid as an element of her password.
ing PassfacesTM (compared to textual passwords),
Images are unique and do not appear more than
even given significant periods of inactivity between
once for a given user. During the authentication
logins.
phase, the same sets of images are shown to the
user, but with the images randomly permuted.
Other studies, e.g., [34, 4], have explored memora-
bility of other types of graphical passwords. We em-
In the Story scheme, a password is a sequence of
phasize, however, that memorability is a secondary
k unique images selected by the user to make a
consideration for our purposes. Our primary goal is
“story”, from a single set of n > k images, each de-
to quantify the effect of user choice on the security
rived from a distinct category of image types. The
of passwords chosen.
images are drawn from categories that depict every-
day objects, food, automobiles, animals, children,
sports, scenic locations, and male and female mod-
els. A sample set of images for the story scheme is
3 Graphical Password Schemes shown in Figure 2.

As mentioned earlier, our evaluation is based on two 3.1 Images
graphical schemes. In the Face scheme, the pass-
word is a collection of k faces, each selected from
a distinct set of n > 1 faces. Each of the n faces As indicated above, the images in each scheme were
are chosen uniformly at random from a set of faces classified into non-overlapping categories. In Face,
classified as belonging to either a “typical” Asian, there were twelve categories: typical Asian males,

Images of “female models” were gathered from a
myriad of pageant sites including Miss USATM , Miss
UniverseTM , Miss NY Chinese, and fashion mod-
eling sites. Images of “male models” were gath-
ered from various online modeling sources including
FordModels.com and StormModels.com.

For the Story scheme, the “men” and “women” cat-
egories were the same as the male and female models
in our Face experiment. All other images were cho-
sen from PicturesOf.NET and span the previously
mentioned categories.

To lessen the effect that an image’s intensity, hue,
and background color may have on influencing a
user choice, we used the ImageMagick library (see
www.imagemagick.org) to set image backgrounds
to a light pastel color at reduced intensity. Ad-
ditionally, images with bright or distracting back-
grounds, or of low quality, were deleted. All remain-
Figure 2: In the Story scheme, a user’s password is ing images were resized to have similar aspect ratios.
sequence of k unique images selected from one set of Of course, it is always possible that differences in
n images, shown above, to depict a “story”. Here, such secondary factors influenced the results of our
n = 9, and images are placed randomly in a 3 × 3 experiment, though we went to significant effort to
grid. avoid this and have found little to support a hypoth-
esis of such influence.

typical Asian females, typical black males, typical 3.2 Experiment
black females, typical white males, typical white
females, Asian male models, Asian female mod-
els, black male models, black female models, white For our empirical evaluation we analyze observa-
male models and white female models. In the Story tions collected during the fall semester (roughly the
scheme, there were nine categories: animals, cars, four month period of late-August through early-
women, food, children, men, objects, nature, and December) of 2003, of graphical password usage by
sports. three separate computer engineering and computer
science classes at two different universities, yielding
The images used for each category were carefully a total of 154 subjects. Each student was randomly
selected from a number of sources. “Typical male” assigned to one of the two graphical schemes. Each
and “typical female” subjects include faces selected student then used the graphical password scheme
from (i) the Asian face database [26] which con- for access to published content including his or
tains color frontal face images of 103 people and her grades, homework, homework solutions, course
(ii) the AR Face database [17] which contains well reading materials, etc., via standard Java enabled
over 4000 color images corresponding to 126 peo- browsers. Our system was designed so that instruc-
ple. For the AR database we used images in angle 2 tors would not post documents on the login server,
only, i.e, frontal images in the smile position. These but rather that this server was merely used to en-
databases were collected under controlled conditions crypt and decrypt documents for posting or retrieval
and are made public primarily for use in evaluating elsewhere. As such, from a student’s perspective,
face recognition technologies. For the most part, the login server provided the means to decrypt doc-
the subjects in these databases are students, and uments retrieved from their usual course web pages.
we believe provide a good representative population
for our study. Additional images for typical male Since there was no requirement for users to change
subjects were derived from a random sampling of their passwords, most users kept one password for
images from the Sports IllustratedTM NBA gallery. the entire semester. However, a total of 174 pass-

Population Scheme 4 Security evaluation
Gender Race Face Story
any any 79 95
Male any 55 77 Recall that in both the Face and Story schemes,
Female any 20 13 images are grouped into non-overlapping categories.
Male Asian 24 27 In our derivations below, we make the simplifying
Female Asian 12 8 assumption that images in a category are equiva-
Male Black 3 - lent, that is, the specific images in a category that
Female Black - - are available do not significantly influence a user’s
choice in picking a specific category.
Male Hispanic - 2
Female Hispanic - -
First we introduce some notation. An -element tu-
Male White 27 48
ple x is denoted x( ) . If S is either the Face or Story
Female White 8 4
scheme, then the expression x( ) ← S denotes the
Table 1: Population breakdown (in passwords). selection of an -tuple x( ) (a password or password
prefix, consisting of image categories) according
to S, involving both user choices and random algo-
rithm choices.

4.1 Password distribution
words were chosen during the semester, implying
that a few users changed their password at least In this section we describe how we approximately
once. During the evaluation period there were a to- compute Pr p(k) ← S for any p(k) , i.e., the proba-
tal of 2648 login attempts, of which 2271 (85.76%)
bility that the scheme yields the password p(k) . This
were successful. Toward the end of the semester,
probability is taken with respect to both random
students were asked to complete an exit survey in
choices by the password selection algorithm and user
which they described why they picked the faces they
choices.
did (for Face) or their chosen stories (for Story)
and provide some demographic information about
We compute this probability inductively as follows.
themselves. This information was used to validate
Suppose p( +1) = q ( ) r(1) . Then
some of our findings which we discuss shortly. Ta-
ble 1 summarizes the demographic information for Pr p( +1)
←S
our users. A gender or race of any includes those for
which the user did not specify their gender or race. = Pr q ( ) ← S ·
Such users account for differences between the sum
of numbers of passwords for individual populations Pr q ( ) r(1) ← S | q ( ) ← S (1)
and populations permitting a race or gender of any.
if p( +1) is valid for S and zero otherwise, where
Pr q (0) ← S = 1. Here, p( +1) is valid iff < k
def
The students participating in this study did so vol-
untarily and with the knowledge they were par- and, for the Story scheme, p( +1) does not con-
ticipating in a study, as required by the Institu- tain any category more than once. The second
tional Review Boards of the participating univer- factor Pr q ( ) r(1) ← S | q ( ) ← S should be under-
sities. However, they were not instructed as to the stood to mean the probability that the user selects
particular factors being studied and, in particular, r(1) after having already selected q ( ) according to
that the passwords they selected were of primary scheme S. If the dataset contains sufficiently many
interest. Nor were they informed of the questions observations, then this can be approximated by
they would be asked at the end of the study. As
such, we do not believe that knowledge of our study # q ( ) r(1) ← S
Pr q ( ) r(1) ← S | q ( ) ← S ≈ ,
influenced their password choices. In addition, since # q( ) ← S
personal information such as their individual grades (2)
were protected using their passwords, we have rea- i.e., using the maximum likelihood estimation,
son to believe that they did not choose them inten- where # x( ) ← S denotes the number of occur-
tionally to be easily guessable. rences of x( ) ← S in our dataset, and where

# x(0) ← S is defined to be the number of pass- Assumption 4.1 permits us to replace (2) by
words for scheme S in our dataset.
Pr ˆ q ( ) r(1) ← S | q ( ) ← S
A necessary condition for the denominator of (2)
ˆ
to be nonzero for every possible q (k−1) is that the # . . . s( ) r(1) ← S
dataset contain N k−1 samples for scheme S where ≈ (4)
N ≥ n denotes the number of image categories for # . . . s( ˆ) ← S
S. (N = 12 in Face, and N = 9 in Story.) N k−1 is
ˆ
over 1700 in the Face scheme, for example. And, of where s( ) is the ˆ-length suffix of q ( ) and we define
course, to use (2) directly to perform a meaningful # . . . s(0) ← S to be the total number of category
approximation, significantly more samples would be choices (k times the number of passwords) in our
required. Thus, we introduce a simplifying, Markov dataset for scheme S. Here, the necessary condition
assumption: a user’s next decision is influenced only for the denominator of (4) to be nonzero for each
ˆ ˆ
by her immediately prior decision(s) (e.g., see [16]). s( ) is that the dataset for S contain N samples,
In other words, rather than condition on all of the e.g., in the Face scheme, twelve for ˆ = 1, and so
previous choices made in a password (q ( ) ), only on.
the last few choices are taken into account. Let
. . . x( ) ← S denote the selection of an -tuple, We further augment the above approach with
≥ , for which the most recent selections are smoothing in order to compensate for gaps in the
x( ) . data (c.f., [16]). Specifically, we replace (4) with

Pr ˆ q ( ) r(1) ← S | q ( ) ← S
Assumption 4.1 There exists a constant ˆ ≥ 0 ˆ
such that if ≥ ˆ then # . . . s( ) r(1) ← S + λ ˆ · Ψ ˆ−1
≈ (5)
# . . . s( ˆ) ← S + λ ˆ
Pr q ( ) r(1) ← S | q ( ) ← S
ˆ
ˆ ˆ
≈ Pr . . . s( ) r(1) ← S | . . . s( ) ← S (3) where s( ) is the ˆ-length suffix of q ( ) ; λ ˆ > 0 is a
real-valued parameter; and where if ˆ > 0 then
ˆ
where s( ) is the ˆ-length suffix of q ( ) . We denote
probabilities under this assumption by Pr ˆ[·]. Ψ ˆ−1 = Pr ˆ−1 q ( ) r(1) ← S | q ( ) ← S

and Ψ ˆ−1 = 1/N otherwise. Note that as λ ˆ is re-
In other words, we assume that if ≥ ˆ, then the duced toward 0, (5) converges toward (4). And,
user’s next selection r(1) is influenced only by her as λ ˆ is increased, (5) converges toward Ψ ˆ−1 , i.e.,
last ˆ choices. This appears to be a reasonable as- a probability under Assumption 4.1 for ˆ − 1, a
sumption, which is anecdotally supported by certain stronger assumption. So, with sufficient data, we
survey answers, such as the following from a user of can use a small λ ˆ and thus a weaker assumption.
the Face scheme. Otherwise, using a small λ ˆ risks relying too heavily
ˆ
on a small number of occurrences of . . . s( ) ← S,
and so we use a large λ ˆ and thus the stronger as-
“To start, I chose a face that stood out from
sumption.
the group, and then I picked the closest face
that seemed to match.”

4.2 Measures
While this user’s intention may have been to choose
a selection similar to the first image she selected, we
conjecture that the most recent image she selected, We are primarily concerned with measuring the abil-
being most freshly on her mind, influenced her next ity of an attacker to guess the password of a user.
choice at least as much as the first one did. Assump- Given accurate values for Pr p(k) ← S for each
tion 4.1 also seems reasonable for the Story scheme p(k) , a measure that indicates this ability is the
on the whole, since users who selected passwords by “guessing entropy” [18] of passwords. Informally,
choosing a story were presumably trying to continue guessing entropy measures the expected number of
a story based on what they previously selected. guesses an attacker with perfect knowledge of the

2^12
probability distribution on passwords would need in
order to guess a password chosen from that distri- 2^10
bution. If we enumerate passwords p1 (k) , p2 (k) , . . .
in non-increasing order of Pr pi (k) ← S , then the 2^8

guessing entropy is simply 2^6

i · Pr pi (k) ← S (6) 2^4

i>0
2^2
Guessing entropy is closely related to Shannon en-
tropy, and relations between the two are known.1 2^0
2^-15 2^-12 2^-9 2^-6 2^-3 2^0 2^3 2^6 2^9 2^12 2^15
Since guessing entropy intuitively corresponds more
!0
closely to the attacker’s task in which we are inter- GS
avg
GSme GS25 GS10 Guessing Entropy

ested (guessing a password), we will mainly consider
measures motivated by the guessing entropy. Figure 4: Measures versus λ0 for Story

The direct use of (6) to compute guessing entropy
using the probabilities in (5) is problematic for two and so we believe this to be a more robust use of
reasons. First, an attacker guessing passwords will (5). We use this sequence to conduct tests with
be offered additional information when performing our dataset in which we randomly select a small
a guess, such as the set of available categories from set of “test” passwords from our dataset (20% of
which the next image can be chosen. For example, the dataset), and use the remainder of the data to
in Face, each image choice is taken from nine images compute the enumeration Π.
that represent nine categories of images, chosen uni-
formly at random from the twelve categories. This We then guess passwords in order of Π until each
additional information constrains the set of possible test password is guessed. To account for the first
passwords, and the attacker would have this infor- issue identified above, namely the set of available
mation when performing a guess in many scenarios. categories during password selection, we first filter
Second, we have found that the absolute probabil- from Π the passwords that would have been invalid
ities yielded by (5) can be somewhat sensitive to given the available categories when the test pass-
the choice of λ ˆ, which introduces uncertainty into word was chosen, and obviously do not guess them.
calculations that utilize these probabilities numeri- By repeating this test with non-overlapping test sets
cally. of passwords, we obtain a number of guesses per
test password. We use Gavg to denote the average
S
2^12
over all test passwords, and Gmed to denote the me-
S
2^10 dian over all test passwords. Finally, we use Gx S
for 0 < x ≤ 100 to denote the number of guesses
2^8
sufficient to guess x percent of the test passwords.
2^6 For example, if 25% of the test passwords could be
guessed in 6 or fewer guesses, then G25 = 6.
S
2^4

2^2
We emphasize that by computing our measures in
this fashion, they are intrinsically conservative given
2^0 our dataset. That is, an attacker who was given 80%
2^-15 2^-12 2^-9 2^-6 2^-3 2^0 2^3 2^6 2^9 2^12 2^15
of our dataset and challenged to guess the remain-
!0
GS
avg
GS
me
GS
25
GS
10 Guessing Entropy ing 20% would do at least as well as our measures
suggest.
Figure 3: Measures versus λ0 for Face

To account for the second of these issues, we use the 4.3 Empirical results
probabilities computed with (5) only to determine
an enumeration Π = (p1 (k) , p2 (k) , . . .) of passwords
in non-increasing order of probability (as computed To affirm our methodology of using Gavg , Gmed , and
S S
with (5)). This enumeration is far less sensitive to Gx as mostly stable measures of password quality,
S
variations in λ ˆ than the numeric probabilities are, we first plot these measures under various instances

of Assumption 4.1, i.e., for various values of ˆ and,
for each, a range of values for λ ˆ. For example, in 2^6.5

the case of ˆ = 0, Figures 3 and 4 show measures
Gavg , Gmed , G25 and G10 , as well as the guessing
S S S S 2^6.0
entropy as computed in (6), for various values of
λ0 . Figure 3 is for the Face scheme, and Figures 4 2^5.5
is for the Story scheme.
2^5.0
The key point to notice is that each of Gavg , Gmed ,
S S
G25 and G10 is very stable as a function of λ0 ,
S S
whereas guessing entropy varies more (particularly 2^4.5 2^10
2^5
for Face). We highlight this fact to reiterate our
2^0
reasons for adopting Gavg , Gmed , and Gx as our
S S S 2^4.0 !0
2^-5
measures of security, and to set aside concerns over 2^-10
2^-5
2^0 2^-10
whether particular choices of λ0 have heavily influ- 2^5
2^10
enced our results. Indeed, even for ˆ = 1 (with some !1

degree of back-off to ˆ = 0 as prescribed by (5)), val-
Figure 6: G25 versus λ0 , λ1 for Face
S
ues of λ0 and λ1 do not greatly impact our measures.
For example, Figures 5 and 6 show Gavg and G25 for
S
S Population Gavg
S Gmed
S G25
S G10
S
Face. While these surfaces may suggest more vari-
ation, we draw the reader’s attention to the small Overall 790 428 112 35
range on the vertical axis in Figure 5; in fact, the Male 826 404 87 53
variation is between only 1361 and 1574. This is in Female 989 723 125 98
contrast to guessing entropy as computed with (6), White Male 844 394 146 76
which varies between 252 and 3191 when λ0 and λ1 Asian Male 877 589 155 20
are varied (not shown). Similarly, while G25 varies
S
between 24 and 72 (Figure 6), the analogous compu- Table 2: Results for Story, λ0 = 2−2
tation using (5) more directly—i.e., computing the
j
smallest j such that i=1 Pr pi
(k)
← S ≥ .25—
varies between 27 and 1531. In the remainder of and the Face scheme, respectively. Populations with
the paper, the numbers we report for Gavg , Gmed ,
S S less than ten passwords are excluded from these ta-
and Gx reflect values of λ0 and λ1 that simultane-
S bles. These numbers were computed under Assump-
ously minimize these values to the extent possible. tion 4.1 for ˆ = 0 in the case of Story and for ˆ = 1 in
the case of Face. λ0 and λ1 were tuned as indicated
in the table captions. These choices were dictated
2^10.65 by our goal of minimizing the various measures we
consider (Gavg , Gmed , G25 and G10 ), though as al-
S S S S
2^10.60 ready demonstrated, these values are generally not
particularly sensitive to choices of λ0 and λ1 .
2^10.55

The numbers in these tables should be considered
2^10.50
in light of the number of available passwords. Story
2^10.45

2^10
Population Gavg
S Gmed
S G25
S G10
S
2^10.40
2^5 Overall 1374 469 13 2
2^0
2^10.35 !0 Male 1234 218 8 2
2^-5
2^-10
2^-5
Female 2051 1454 255 12
2^0 2^-10
2^5 Asian Male 1084 257 21 5.5
2^10
!1 Asian Female 973 445 19 5.2
White Male 1260 81 8 1.6
Figure 5: Gavg versus λ0 , λ1 for Face
S
Table 3: Results for Face, λ0 = 2−2 , λ1 = 22
Tables 2 and 3 present results for the Story scheme

has 9 × 8 × 7 × 6 = 3024 possible passwords, yielding Moreover, there was also significant correlation
a maximum possible guessing entropy of 1513. Face, among members of the same race. As shown in Ta-
on the other hand, has 94 = 6561 possible passwords ble 5, Asian females and white females chose from
(for fixed sets of available images), for a maximum within their race roughly 50% of the time; white
guessing entropy of 3281. males chose whites over 60% of the time, and black
males chose blacks roughly 90% of the time (though
Our results show that for Face, if the user is known the reader should be warned that there were only
to be a male, then the worst 10% of passwords can three black males in the study, thus this number re-
be easily guessed on the first or second attempt. quires greater validation). Again, a number of exit
This observation is sufficiently surprising as to war- surveys confirmed this correlation, e.g.:
rant restatement: An online dictionary attack of
passwords will succeed in merely two guesses for
“I picked her because she was female and
10% of male users. Similarly, if the user is Asian Asian and being female and Asian, I thought
and his/her gender is known, then the worst 10% of I could remember that.”
passwords can be guessed within the first six tries.
“I started by deciding to choose faces of
It is interesting to note that Gavg is always higher
S people in my own race ... specifically, people
than Gmed . This implies that for both schemes,
S that looked at least a little like me. The
there are several good passwords chosen that sig- hope was that knowing this general piece of
nificantly increase the average number of guesses information about all of the images in my
an attacker would need to perform, but do not af- password would make the individual faces
fect the median. The most dramatic example of easier to remember.”
this is for white males using the Face scheme, where
Gavg = 1260 whereas Gmed = 81.
S S “... Plus he is African-American like me.”

These results raise the question of what different Female Male Typical Typical
populations tend to choose as their passwords. In- Pop. Model Model Female Male
sight into this for the Face scheme is shown in Ta- Female 40.0% 20.0% 28.8% 11.3%
bles 4 and 5, which characterize selections by gender Male 63.2% 10.0% 12.7% 14.0%
and race, respectively. As can be seen in Table 4,
both males and females chose females in Face signif- Table 4: Gender and attractiveness selection in
icantly more often than males (over 68% for females Face.
and over 75% for males), and when males chose fe-
males, they almost always chose models (roughly
80% of the time). These observations are also widely Insight into what categories of images different gen-
supported by users’ remarks in the exit survey, e.g.: ders and races chose in the Story scheme are shown
in Tables 6 and 7. The most significant deviations
between males and females (Table 6) is that females
“I chose the images of the ladies which chose animals twice as often as males did, and males
appealed the most.”
chose women twice as often as females did. Less
pronounced differences are that males tended to se-
“I simply picked the best lookin girl on each lect nature and sports images somewhat more than
page.” females did, while females tended to select food im-
ages more often. However, since these differences
“In order to remember all the pictures for my
login (after forgetting my ‘password’ 4 times
in a row) I needed to pick pictures I could Pop. Asian Black White
EASILY remember - kind of the same pitfalls Asian Female 52.1% 16.7% 31.3%
when picking a lettered password. So I chose Asian Male 34.4% 21.9% 43.8%
all pictures of beautiful women. The other Black Male 8.3% 91.7% 0.0%
option I would have chosen was handsome White Female 18.8% 31.3% 50.0%
men, but the women are much more pleasing
White Male 17.6% 20.4% 62.0%
to look at :)”
Table 5: Race selection in Face.
“Best looking person among the choices.”

100%
were all within four percentage points, it is not clear Face Story
98%
how significant they are. Little emerges as definitive
96%
trends by race in the Story scheme (Table 7), par-
94%
ticularly considering that the Hispanic data reflects

Correct Login %
92%
only two users and so should be discounted.
90%

88%

86%

84%
5 Memorability evaluation 82%

80%
0 4 8 12 16
In this section we briefly evaluate the memorabil- Time Since Last Login Attempt (days)

ity of the schemes we considered. As described in
Section 2, there have been many usability studies Figure 8: Memorability versus time since last login
performed for various graphical password schemes, attempt. Each data point represents the average of
including for variants of the Face scheme. As such, 90 login attempts.
our goal in this section is not to exhaustively eval-
uate memorability for Face, but rather to simply
benchmark the memorability of the Story scheme ently few of them actually chose stories, despite our
against that of Face to provide a qualitative and suggestion to do so. Nearly 50% of Story users re-
relative comparison between the two. ported choosing no story whatsoever in their exit
surveys. Rather, these users employed a variety of
Figure 7 shows the percentage of successful logins alternative strategies, such as picking four pleasing
versus the amount of time since the password was pictures and then trying to memorize the order in
initially established, and Figure 8 shows the per- which they picked them. Not surprisingly, this con-
centage of successful logins versus the time since tributed very significantly to incorrect password en-
that user’s last login attempt. Each figure includes tries due to misordering their selections. For exam-
one plot for Face and one plot for Story. A trend ple, of the 236 incorrect password entries in Story,
that emerges is that while memorability of both over 75% of them consisted of the correct images se-
schemes is strong, Story passwords appear to be lected in an incorrect order. This is also supported
somewhat harder to remember than Face. We do anecdotally by several of the exit surveys:
not find this to be surprising, since previous studies
have shown Face to have a high degree of memora- “I had no problem remembering the four
bility. pictures, but I could not remember the
100%
original order.”
Face Story

95% “No story, though having one may have helped
to remember the order of the pictures better.”
Correct Login %

90%

85% “... but the third try I found a sequence
that I could remember. fish-woman-girl-corn,
80% I would screw up the fish and corn order 50%
of the time, but I knew they were the pic-
75% tures.”

70%
0 20 40 60 80
Time Since Password Change (days) As such, it seems advisable in constructing graphical
password schemes to avoid having users remember
Figure 7: Memorability versus time since password an ordering of images. For example, we expect that
change. Each data point represents the average of a selection of k images, each from a distinct set of
100 login attempts. n images (as in the Face scheme, though with im-
age categories not necessarily of only persons), will
One potential reason for users’ relative difficulty in generally be more memorable than an ordered se-
remembering their Story passwords is that appar- lection of k images from one set. If a scheme does

Pop. Animals Cars Women Food Children Men Objects Nature Sports
Female 20.8% 14.6% 6.3% 14.6% 8.3% 4.2% 12.5% 14.6% 4.2%
Male 10.4% 17.9% 13.6% 11.0% 6.8% 4.6% 11.0% 17.2% 7.5%

Table 6: Category selection by gender in Story

Pop. Animals Cars Women Food Children Men Nature Objects Sports
Asian 10.7% 18.6% 11.4% 11.4% 8.6% 4.3% 17.1% 11.4% 6.4%
Hispanic 12.5% 12.5% 25.0% 12.5% 0.0% 12.5% 12.5% 12.5% 0.0%
White 12.5% 16.8% 13.0% 11.5% 6.3% 4.3% 16.8% 11.1% 7.7%

Table 7: Category selection by race in Story

rely on users remembering an ordering, then the im- (as is [4]), and our study indicates that password se-
portance of the story should be reiterated to users, lection in this scheme is sufficiently free from bias to
since if the sequence of images has some semantic suggest that reasonable limits could be imposed on
meaning then it is more likely that the password password guesses to render the scheme secure. For
is memorable (assuming that the sequences are not example, the worst 10% of passwords in the Story
too long [21]). scheme for the most predictable population (Asian
males) still required twenty guesses to break, sug-
gesting a limit of five incorrect password guesses
might be reasonable, provided that some user ed-
6 Conclusion ucation is also performed.

The relative strength of the Story scheme must be
The graphical password schemes we considered in balanced against what appears to be some difficulty
this study have the property that the space of pass- of memorability for users who eschew the advice of
words can be exhaustively searched in short order using a story to guide their image selection. An al-
if an offline search is possible. So, any use of these ternative (besides better user education) is to per-
schemes requires that guesses be mediated and con- mit unordered selection of images from a larger set
firmed by a trusted online system. In such scenarios, (c.f., [4, 7]). However, we believe that further, more
we believe that our study is the first to quantify fac- sizeable studies must be performed in order to con-
tors relevant to the security of user-chosen graphical firm the usability and security of these approaches.
passwords. In particular, our study advises against
the use of a PassfacesTM -like system that permits
user choice of the password, without some means to
mitigate the dramatic effects of attraction and race 7 Acknowledgments
that our study quantifies. As already demonstrated,
for certain populations of users, no imposed limit
on the number of incorrect password guesses would The authors would like to thank Joanne Houlahan
suffice to render the system adequately secure since, for her support and for encouraging her students
e.g., 10% of the passwords of males could have been to use the graphical login server. We also extend
guessed by merely two guesses. our gratitude to all the students at Carnegie Mel-
lon University and Johns Hopkins University who
Alternatives for mitigating this threat are to pro- participated in this study.
hibit or limit user choice of passwords, to educate
users on better approaches to select passwords, or to
select images less prone to these types of biases. The
first two are approaches initially attempted in the Notes
context of text passwords, and that have appeared
in some graphical password schemes, as well. The
Story scheme is one example of the third strategy

1 For a random variable X taking on values in X , if G(X) [10] J. Langlois, L. Kalakanis, A. Rubenstein,
denotes its guessing entropy and H(X) denotes its Shannon A. Larson, M. Hallam, and M. Smoot. Max-
entropy, then it is known that G(X) ≥ 2H(X)−2 + 1 [18] and
2 log |X | ims and myths of beauty: A meta-analytic
that H(X) ≥ |X |−1 (G(X) − 1) [19].
and theoretical review. In Psychological Bul-
letin 126:390–423, 2000.

[11] D. Levin. Race as a visual feature: using visual
References search and perceptual discrimination tasks to
understand face categories and the cross race
recognition deficit. Quarterly Journal of Exper-
[1] T. Alley and M. Cunningham. Averaged faces
imental Psychology:General, 129 (4), 559-574.
are attractive, but very attractive faces are not
average. In Psychological Science, 2, pages 123- [12] D. Lindsay, P. Jack, and M. Chrisitan. Other-
125, 1991. race face perception. Journal of Applied Psy-
chology 76:587–589, 1991.
[2] G. E. Blonder. Graphical password. US Patent
5559961, Lucent Technologies, Inc., Murray [13] T. Luce. Blacks, whites and yellows: They all
Hill, NJ, August 30, 1995. look alike to me. Psychology Today 8:105–108,
1974.
[3] S. Brostoff and M.. A. Sasse. Are PassfacesTM
more usable than passwords? A field trial [14] S. Madigan. Picture memory. In Imagery,
investigation. In Proceedings of Human Com- Memory, and Cognition, pages 65–86,
puter Interaction, pages 405–424, 2000. Lawrence Erlbaum Associates, 1983.

[4] R. Dhamija and A. Perrig. D´j` vu: A user
ea [15] R. S. Malpass. They all look alike to me. In The
study using images for authentication. In Pro- Undaunted Psychologist, pages 74-88, McGraw-
ceedings of the 9th USENIX Security Sympo- Hill, 1992.
sium, August 2000.
[16] C. Manning and H. Sch¨tze. Foundations of
u
[5] D. Feldmeier and P. Karn. UNIX password Statistical Natural Language Processing, Chap-
security—Ten years later. In Advances in ter 6, MIT Press, May 1999.
Cryptology—CRYPTO ’89 (Lecture Notes in
Computer Science 435), 1990. [17] A. M. Martinez and R. Benavente. The AR
Face Database. Technical Report number 24,
[6] A. Feingold. Good-looking people are not what June, 1998.
we think. In Psychological Bulletin, 111, pages
[18] J. L. Massey. Guessing and entropy. In Pro-
304-341, 1992.
ceedings of the 1994 IEEE International Sym-
[7] W. Jansen, S. Gavrila, V. Korolev, R. Ayers, posium on Information Theory, 1994.
and R. Swanstrom. Picture password: A
[19] R. J. McEliece and Z. Yu. An inequality on
visual login technique for mobile devices.
entropy. In Proceedings of the 1995 IEEE In-
NISTIR 7030, Computer Security Division,
ternational Symposium on Information Theory,
Information Technology Laboratory, Na-
1995.
tional Institute of Standards and Technology,
Gaithersburg, MD, July 2003. Available [20] C. Meissner, J. Brigham. Thirty years of in-
at http://csrc.nist.gov/publications/ vestigation the own-race advantage in memory
nistir/nistir-7030.pdf. for faces: A meta-analytic review. Psychology,
Public Policy & Law, 7, pages 3-35, 2001.
[8] I. Jermyn, A. Mayer, F. Monrose, M. Re-
iter and A. Rubin. The design and analysis of [21] G. A. Miller. The magical number seven, plus
graphical passwords. In Proceedings of the 8th or minus two: Some limits on our capacity
USENIX Security Symposium, August 1999. for processing information. Psychological Re-
view 63:81–97, 1956.
[9] D. Klein. Foiling the cracker: A survey of, and
improvements to, password security. In Pro- [22] R. Morris and K. Thompson. Password secu-
ceedings of the 2nd USENIX Security Work- rity: A case history. Communications of the
shop, pages 5–14, August 1990. ACM 22(11):594–597, November 1979.

[23] D. L. Nelson, U. S. Reed, and J. R. Walling.
Picture superiority eﬀect. Journal of Exper-
imental Psychology: Human Learning and
Memory, 3:485–497, 1977.

[24] The Science Behind Passfaces. Revision 2,
Real User Corporation, September 2001.
Available at http://www.realuser.com/
published/ScienceBehindPassfaces.pdf.

[25] Strategies for using PassfacesTM for Win-
dows. Real User Corporation, 2002. Available
at http://www.realuser.com/published/
PassfacesforWindowsStrategies.pdf.

[26] Asian Face Image Database PF01. Pohang Uni-
versity of Science and Technology, Korea, 2001.

[27] T. Valentine. An evaluation of the PassfacesTM
personal authentication system. Technical Re-
port, Goldsmiths College University of London,
1998.

[28] T. Valentine. Memory for PassfacesTM after a
long delay. Technical Report, Goldsmiths Col-
lege University of London, 1999.

[29] T. Valentine and M. Endo. Towards an exem-
plar model of face processing: The eﬀects of
race and distinctiveness. Quarterly Journal of
Experimental Psychology 44, 671-703.

[30] Visual Key – Technology. Available at http:
//www.viskey.com/tech.html.

[31] P. Walker and W. Tanaka. An encoding ad-
vantage for own-race versus other-race faces. In
Perception, 23, pages 1117-1125, 2003.

[32] T. Wu. A real-world analysis of Kerberos pass-
word security. In Proceedings of the 1999 ISOC
Symposium on Network and Distributed System
Security, February 1999.

[33] M. Zviran and W. J. Haga. Cognitive pass-
words: The key to easy access and control.
Computers and Security 9(8):723–736, 1990.

[34] M. Zviran and W. J. Haga. A comparison
of password techniques for multilevel authen-
tication mechanisms. The Computer Journal
36(3):227–237, 1993.

Usenix04

More Related Content

What's hot (20)

Viewers also liked (7)

Similar to Usenix04 (20)

Usenix04