Skip to content

Athena-class round by round Risk Limiting Audit including Minerva

License

Notifications You must be signed in to change notification settings

gwexploratoryaudits/athena

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Athena - Risk Limiting Audit (Round-by-Round)

This library is an implementation of The ATHENA Class of Risk-Limiting Ballot Polling Audits

Installation

python -m pip install git+https://github.com/filipzz/athena.git

Usage

As a library

See athena.ipynb

As a command line

Entering data from command line

python -m athena --name contestName --ballots 5000 3000 2000 --pstop .9 .95 --type athena

where:

  • --ballots is a list of votes each candidate received
  • --name name of a contest
  • --pstop is a list of stopping probabilities
  • --type is an optional argument that lets one see result of different audit types athena (default) | bravo (wald) | arlo

The above call would return:

[[0.9, 148], [0.95, 213]]

which means that if 148 will be drawn the audit would stop with probability 90% while if 213 ballots are drawn then the probability of completing audit would be 95%.

The above call can be and with shorter parameter names:

python -m athena -n contestName -b 5000 3000 2000  -p .9 --type arlo

Reading data from a file

Election data can be read from a json file of the following form:

 {
	"name": "Local Elections 2020",
	"total_ballots": 10000,
	"contests": {
		"contest_1": {
			"contest_ballots": 10000,
			"tally": {
				"A": 6000,
				"B": 4000
			},
			"num_winners": 1,
			"reported_winners": [
				"A"
			],
			"contest_type": "PLURALITY"
		}
    }
}

A file may contain many contests.

Then one needs to provide a path to a file (parameter --file) and provide the name of the contest (with the example above: -n contest_1):

python -m athena --file  athena/test_data/simple.json --new contest_1 --pstop .7

Usage

Athena helps to:

  • estimate the number of ballots to be drawn in order to complete audit with a

Finding stopping probabilities

In this mode, one can (roughly) what are the expected round sizes for a given results:

  • --ballots (list of declared number of votes cast for each candidate)
  • --pstop (list of stopping probabilities for the next round)
python -m athena -n asd -b 60000 40000 --pstop .7 .8 .9 

returns:

[0.7, 0.8, 0.9]
Results of: asd
Number of valid ballots: 100 000
	1 A	60 000
	2 B	40 000

Parameters: 
Alpha:  0.1
Gamma:  1.0
Model:  bin
Round schedule: []
setting round schedule


A (60 000) vs B (40 000)
	margin:	0.2
pstop goal: [0.7, 0.8, 0.9]
round schedule: []
	0.7	[112]
	0.8	[132]
	0.9	[180]

If one want to estimate round sizes for the next round, round_schedule parameter needs to be added

python -m athena -n asd -b 60000 40000 --pstop .5 .8  --rounds 132
[0.5, 0.8]
Results of: asd
Number of valid ballots: 100 000
	1 A	60 000
	2 B	40 000

Parameters: 
Alpha:  0.1
Gamma:  1.0
Model:  bin
Round schedule: [132]
setting round schedule


A (60 000) vs B (40 000)
	margin:	0.2
pstop goal: [0.5, 0.8]
round schedule: [132]
	0.5	[132, 226]
	0.8	[132, 302]

To get obtain detailed information about a selected round schedule (e.g., [132, 297]) one needs to call it without --pstop parameter.

python3 -m athena -n asd -b 60000 40000 --rounds 132 297
Results of: asd
Number of valid ballots: 100000
	1 A	60000
	2 B	40000

Parameters: 
Alpha:  0.1
Gamma:  1.0
Model:  bin
Round schedule: [132, 297]


A (60000) vs B (40000)
	margin:	0.2

	Approx round schedule:	[132, 297]
	ATHENA kmins:		[75, 165]
	ATHENA pstop (audit):	[0.798620073240586, 0.9604763404354486]
	ATHENA pstop (tied): 	[0.06933400321906039, 0.08488117767255865]
	ATHENA gammas ():	[0.3848374142119376, 0.5336003665053669]
	ATHENA true risk:	[0.0868172558419695, 0.08837404327323284]

At the end we may want to check how, for that round schedule Arlo audit would work

python3 -m athena -n asd -b 60000 40000 --rounds 132 297 --type arlo
Results of: asd
Number of valid ballots: 100000
	1 A	60000
	2 B	40000

Parameters: 
Alpha:  0.1
Gamma:  1.0
Model:  bin
Round schedule: [132, 297]


A (60000) vs B (40000)
	margin:	0.2

	Approx round schedule:	[132, 297]
	ARLO kmins:		[79, 170]
	ARLO pstop (audit):	[0.5517707405988953, 0.8681144359578333]
	ARLO pstop (tied): 	[0.014585813499702702, 0.019495722498663794]
	ARLO gammas ():	[0.07601726700482742, 0.07026836102128296]
	ARLO true risk:	[0.02643455411186025, 0.022457549017893253]

Wald's sequential test

Wald's sequential test (BRAVO audit) can be seen as a special case of AURROR. AURROR with round schedule [1, 2, 3, ..., max] is the same as Wald's sequential test that stops after up to max ballots checked.

Calling:

python3 -m athena -n asd -b 6000 4000  --rounds 17 --type BRAVO

Is equivalent to calling:

python3 -m athena -n asd -b 6000 4000  --rounds 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

Note that athena uses the code that is optimized for a few rounds of (potentially) large sizes while BRAVO test has each round of size 1. Therefore you may prefere to run a previous version of the code:

python3 code/aurror_old.py -n asd -b 6000 4000  --rounds 17

Estimating true audit risk

python -m athena -n 2016_Minnesota  -b 1367825 1323232  --rounds 14880 32138   --risk 7530


{'2016_Minnesota': {"round_number": 2, "min_kmins": [7531, 0], "risks": [0.10015171727863798]}}

Help

python -m athena -h

returns:

usage: athena.py [-h] [-v] [-n NEW] [-a ALPHA]
                 [-c [CANDIDATES [CANDIDATES ...]]]
                 [-b [BALLOTS [BALLOTS ...]]] [-t TOTAL]
                 [-r ROUNDS [ROUNDS ...]] [-p PSTOP [PSTOP ...]] [-w WINNERS]
                 [-l LOAD] [--type TYPE] [-e RISK [RISK ...]]

This program lets for computing AURROR parameters.

optional arguments:
  -h, --help            show this help message and exit
  -v, -V, --version     shows program version
  -n NEW, --new NEW     creates new election folder where all data are stored
  -a ALPHA, --alpha ALPHA
                        set alpha (risk limit) for the election
  -c [CANDIDATES [CANDIDATES ...]], --candidates [CANDIDATES [CANDIDATES ...]]
                        set the candidate list (names)
  -b [BALLOTS [BALLOTS ...]], --ballots [BALLOTS [BALLOTS ...]]
                        set the list of ballots cast for every candidate
  -t TOTAL, --total TOTAL
                        set the total number of ballots in given contest
  -r ROUNDS [ROUNDS ...], --rounds ROUNDS [ROUNDS ...], --round_schedule ROUNDS [ROUNDS ...]
                        set the round schedule
  -p PSTOP [PSTOP ...], --pstop PSTOP [PSTOP ...]
                        set stopping probability goals for each round
                        (corresponding round schedule will be found)
  -w WINNERS, --winners WINNERS
                        set number of winners for the given race
  -l LOAD, --load LOAD  set the election to read
  --type TYPE           set the audit type (BRAVO/AURROR)
  -e RISK [RISK ...], --risk RISK [RISK ...], --evaluate_risk RISK [RISK ...]
                        evaluate risk for given audit results

Unit tests

Outside the code directory: pytest-3 code (or python -m pytest code).

Inside the code directory: pytest-3 (or python -m pytest)

Old version BRAVO/ARLO/ATHENA

Previous implementations:

  • code/aurror_old.py
  • code/aurror.py

were here up to the following commit.

python3 aurror_old.py --new raceName --alpha .1 --candidates "Albus D." Bob Cedric --ballots 5000 3000 2000 --round_schedule 100 200
Results of: raceName
Number of valid ballots: 10000
	1 Albus D.	5000
	2 Bob	3000
	3 Cedric	2000

Parameters: 
Alpha:  0.1
Gamma:  1.0
Model:  bin
Round schedule: [100, 200]


Albus D. (5000) vs Bob (3000)
	margin:	0.25

	Effective round schedule: [80, 160]

	BRAVO kmins: 	[50, 95]
	BRAVO risk: 	[0.06235051 0.08209004]
	BRAVO pstop: 	[0.6862955 0.9054928]
	--- ratio:	[0.09085082782119536, 0.09065786273502807]

	ARLO kmins:	[50, 95]
	ARLO risk:	[0.01649631 0.0234236 ]
	ARLO pstop:	[0.54969483 0.83843797]
	--- ratio:	[0.030009940780989588, 0.027937185688492874]

	Athena kmins:	[47, 92]
	Athena risk:	[0.07281773 0.08800107]
	Athena pstop:	[0.79143613 0.93830014]
	--- ratio:	[0.092007080402985, 0.09378776507297898]


Albus D. (5000) vs Cedric (2000)
	margin:	0.42857142857142855

	Effective round schedule: [70, 140]

	BRAVO kmins: 	[46, 89]
	BRAVO risk: 	[0.07876542 0.08264516]
	BRAVO pstop: 	[0.95085346 0.99592736]
	--- ratio:	[0.08283654763927709, 0.08298311619828014]

	ARLO kmins:	[46, 89]
	ARLO risk:	[0.0057632  0.00623018]
	ARLO pstop:	[0.88184075 0.98651178]
	--- ratio:	[0.006535425834876662, 0.006315362816084122]

	Athena kmins:	[41, 83]
	Athena risk:	[0.0941081  0.09882721]
	Athena pstop:	[0.99258982 0.99961812]
	--- ratio:	[0.0948106654177994, 0.09886496664859996]


Bob (3000) vs Cedric (2000)
	margin:	0.2

	Effective round schedule: [50, 100]

	BRAVO kmins: 	[34, 61]
	BRAVO risk: 	[0.02323915 0.05373822]
	BRAVO pstop: 	[0.25028037 0.5834069 ]
	--- ratio:	[0.0928524486015076, 0.09211103820218584]

	ARLO kmins:	[34, 61]
	ARLO risk:	[0.00767334 0.02212395]
	ARLO pstop:	[0.1560906  0.47770637]
	--- ratio:	[0.04915951815417218, 0.04631285308993794]

	Athena kmins:	[32, 59]
	Athena risk:	[0.03245432 0.0627022 ]
	Athena pstop:	[0.33561326 0.65119262]
	--- ratio:	[0.09670155223011023, 0.0962882608728027]

Acknowledgements

We thank NSF (1421373 TWC: TTP Option: Small: Open-Audit Voting Systems---Protocol Models and Properties) for funding the implementation of this project.

About

Athena-class round by round Risk Limiting Audit including Minerva

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 89.1%
  • Jupyter Notebook 10.9%