Next: Measures
Up: Security evaluation
Previous: Security evaluation
Password distribution
In this section we describe how we approximately compute
for any , i.e.,
the probability that the scheme yields the password .
This probability is taken with respect to both random choices by the
password selection algorithm and user choices.
We compute this probability inductively as follows. Suppose
. Then
if
is valid for and zero otherwise, where
. Here,
is valid iff and, for the Story
scheme,
does not contain any category more than
once. The second factor
should be understood to
mean the probability that the user selects after having
already selected according to scheme . If the
dataset contains sufficiently many observations, then this can be
approximated by
|
(2) |
i.e., using the maximum likelihood estimation, where
denotes the number of
occurrences of
in our dataset, and where
is defined to be the number
of passwords for scheme in our dataset.
A necessary condition for the denominator of (2) to be
nonzero for every possible is that the dataset contain
samples for scheme where denotes the number
of image categories for . ( in Face, and in Story.)
is over 1700 in the Face scheme, for example. And, of
course, to use (2) directly to perform a meaningful
approximation, significantly more samples would be required. Thus, we
introduce a simplifying, Markov assumption: a user's next decision is
influenced only by her immediately prior decision(s) (e.g.,
see [16]). In other words, rather than condition on all of the
previous choices made in a password (), only the last
few choices are taken into account. Let
denote the selection of an -tuple,
, for
which the most recent selections are .
Assumption 4.1
There exists a constant
such that if
then
where
is the
-length suffix of
. We denote probabilities under this assumption by
.
In other words, we assume that if
, then the
user's next selection is influenced only by her last
choices. This appears to be a reasonable assumption,
which is anecdotally supported by certain survey answers, such as the
following from a user of the Face scheme.
``To start, I chose a face that stood out from the group, and then I
picked the closest face that seemed to match.''
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, being most freshly on her mind, influenced
her next choice at least as much as the first one did.
Assumption 4.1 also seems reasonable for the Story
scheme on the whole, since users who selected passwords by choosing a
story were presumably trying to continue a story based on what they
previously selected.
Assumption 4.1 permits us to replace
(2) by
where
is the -length suffix of
and we define
to be the total
number of category choices ( times the number of passwords) in our
dataset for scheme . Here, the necessary condition for the
denominator of (4) to be nonzero for each
is that the dataset for contain
samples, e.g., in the Face scheme, twelve
for
, and so on.
We further augment the above approach with smoothing in order to
compensate for gaps in the data (c.f., [16]). Specifically, we
replace (4) with
where
is the -length suffix of
;
is a real-valued parameter; and where
if
then
and
otherwise.
Note that as
is reduced toward ,
(5) converges toward (4). And, as
is increased, (5) converges
toward
, i.e., a probability under
Assumption 4.1 for , a stronger
assumption. So, with sufficient data, we can use a small
and thus a weaker assumption. Otherwise,
using a small
risks relying too heavily
on a small number of occurrences of
,
and so we use a large
and thus the stronger assumption.
Next: Measures
Up: Security evaluation
Previous: Security evaluation