Você está na página 1de 7

How to compute xs genuine and imposter scores

Step 1: Create xs Profile (i.e., summary of characteristics/features )


x(i.e., genuine user) Collect samples Create xs profile from xs samples Individual specific threshold:
If we have sufficient number
Step 2: Computing genuine scores of genuine and impostor
scores for x, then we can
x Collect more samples Compare each of these xs compute individual specific
samples with xs profile (created in threshold for x using the
step 1). For each sample there will threshold
be one genuine score computation method we
Step 2: Computing impostor scores learnt in the past.

Other users Collect impostor Compare each of the impostor For different individuals
(impostors, i.e., samples samples with xs profile (created in we have to compute
not x) step 1). For each impostor sample different individual
there will be one impostor score specific thresholds.
689599.7 rule

% of people

99.85% scores fall


99.85% on the right side of
-3
49.85 % 50%
Similarly we can
show that, 99.85%
scores fall on the
left side of +3
Score
For the normal distribution, the values less than one standard deviation away from the mean
covers 68.27% of all scores; while two standard deviations from the mean covers 95.45% of all
scores; and three standard deviations covers 99.73% of all scores.
3 Method for population based threshold Genuine score for x : User xs profile
These two graphs are based on all versus user xs sample
users genuine and impostor scores
Imposter score for x: User xs profile
versus user ys sample where x and y
are different individuals
Individual specific threshold is an
unique threshold for a user x (custom
sizing) computed using xs genuine
and impostor scores
Population based threshold is a single
threshold for all users (one size for
all) computed using all of their
genuine and impostor scores

If we use -3 as the threshold, 99.85% + classes (genuine scores) will be truly


identified as +. But some classes (impostor scores) may cross -3 and be falsely
identified as +. The following formulation minimizes these errors:

Threshold
Furuis method for computing individual
specific thresholds using just imposter scores
Threshold for x

Where, and are population based constants (i.e., same values for all) ,
and are respectively the mean and standard deviation of impostor
scores for x
This method does not require xs genuine scores, and hence, can be
computed collecting just some profile data from x in short time.

The same set of imposter samples can be used to compute imposter


scores for any user (who didnt volunteer as imposter), i.e., we dont have
to collect imposter samples again and again for every new user.
Assignment 2.
Write a program to Compute threshold for John using Furuis method.
and for John are .48 and .08 respectively. Use the following table to
compute and .
User# Known
Threshold Consider the possible values of and from the
1 .8 . 46 .1 following set:
{0, .1, .2, .3, .4, .5, .6, .7, .9, 1}
2 .65 .38 .08
3 .7 .40 .07 You can use any language to write the program
4 .7 .44 .06
Submit (i) the program/code and (ii) Johns
5 .75 .5 .09
computed threshold
6 .72 .6 .01
7 .76 .46 .04
8 .78 .48 .08
9 .77 .52 .09
10 .75 .45 .12
Assignment 2: Hints find and that minimizes
RMS of errors
User# Known Computed threshold for Error RMS of errors
Threshold a given pair of and (Th-Th)
(Th) (Th)
1
2
3
4 Square root of
5 the summation
of the squares of
6 individual errors
7
8
9
10
Why we use RMS of errors?
How can we determine which of the following error sets is better?

A={User 1 error =.01, User 2 error =.2, User 3 error = .15}


versus
B ={User 1 error =.13, User 2 error =.19, User 3 error = .14}

.01 is way better than .13


.2 is slightly worse than .19
.15 is slightly worse than .14

So we cannot directly compare. If we compute RMS of A and B, then we can compare.

Você também pode gostar