SLIDE 1 So#ware
Support
for
So#ware‐Independent Audi3ng
Gabrielle
A.
Gianelli,
Jennifer
D.
King, Edward
W.
Felten,
William
P.
Zeller
Center
for
Informa3on
Technology
Policy Department
of
Computer
Science Princeton
University
EVT/WOTE
‘09 August
11,
2009
SLIDE 2 Goals
of
Post‐Elec3on
Audi3ng
- Valid
sta3s3cal
guarantee
- Efficient
- Easy
to
use
- Ins3lls
confidence
in
elec3on
results
for
voters
and
officials
SLIDE 3
SLIDE 4
Possible
Solu3ons
Efficiency Complexity Difficulty
understanding
SLIDE 5
Possible
Solu3ons
Ease
of
use Automa9on So;ware
dependence
?
SLIDE 6 How
to
reconcile
these
tensions?
- One
approach:
eliminate
computers
- Our
approach:
automate,
but
verify
SLIDE 7
So#ware
Independence
Audi9ng System Third
Par3es for Verifica3on
SLIDE 8
Log
Format
User
entered
data Calcula9ons ‐
Pseudorandom
numbers ‐
Input ‐
Precinct
or
ballot
selec3ons
SLIDE 9
Log
AXributes
of
log
– XML:
can
be
easily
parsed – Stores
all
informa3on
necessary
to
recreate
an
audit, either
by
hand
or
with
another
machine
A log verifiable by a third party ensures software independence.
SLIDE 10 Our
Solu3on
- Web
applica3on
- Python
with
Django
web
framework
Goal: auditing interface easy for non-expert users
SLIDE 11
- (screenshot
of
home
page
with
audit
status)
SLIDE 12
- (screenshot
of
race/algorithm
selec3on
page)
SLIDE 13 Supported
Algorithms
Precinct‐based
algorithms:
- Exact
Percent
- Percent
by
Probability
Ballot‐based
algorithms:
- Constant
Sample
Size
- Varying
Sample
Size
SLIDE 14 Linking
Precincts
- Assume
two
races
A
and
B
over
the
same
set
of
precincts
- Goal:
choose
2%
of
precincts
for
Race
A
and
3%
of
precincts
for
Race
B
SLIDE 15
Unlinked
Precincts
Set
of
all
ballots,
S SB
SA
SLIDE 16
Linked
Precincts
Set
of
all
ballots,
S
SB SA&
SB
SLIDE 17
Pseudorandom
Number
Genera3on
PRNG “1,2,1,4,4,…”
SLIDE 18 Humboldt
County
Data
- Ballot
images
from
Humboldt
County
(CA)
Elec3on
Transparency
Project
(Nov
2008)
- Textual
ballot
representa3ons
from
Mitch
Trachtenberg’s
Ballot
Browser
program
- 29
races;
145
precincts;
128,144
ballots
SLIDE 19 Process
- Loaded
the
data
from
individual
ballots
into
our
database
- Used
the
system
to
run
a
mock
audit
- In
order
to
simulate
a
manual
recount,
compared
the
ballot
images
against
the
data
in
our
database
SLIDE 20
Results
2% 4% 12% Percent
ballots chosen 3,006 N/A 99% confidence Constant Sample
Size 3 5,768 15 1%
of precincts Exact
Percent (linking) 2 15,613 33 1%
of precincts Exact
Percent 1 Ballots Chosen Precincts chosen Parameter Algorithm Audit
SLIDE 21 In
closing…
Automa3on
can
- Make
post‐elec3on
audits
more
efficient
- Expand
the
scope
of
complex
audi3ng
algorithms
and
reduce
the
number
of
ballots
to
be
counted as
long
as
the
output
can
be
independently
verified.
SLIDE 22 So#ware
Support
for
So#ware‐Independent Audi3ng
Gabrielle
A.
Gianelli,
Jennifer
D.
King, Edward
W.
Felten,
William
P.
Zeller
Center
for
Informa3on
Technology
Policy Department
of
Computer
Science Princeton
University
EVT/WOTE
‘09 August
11,
2009
