SLIDE 1
Weight, Weight, Don’t Tell Me: Using Scales to Select Ballots for Audi<ng
Cynthia Sturton, Eric Rescorla, David Wagner
1
Elec<on Audits are Important
2
Source: Joe Hall
Weight, Weight, Dont Tell Me: Using Scales to Select Ballots for - - PowerPoint PPT Presentation
Weight, Weight, Dont Tell Me: Using Scales to Select Ballots for - - PowerPoint PPT Presentation
Weight, Weight, Dont Tell Me: Using Scales to Select Ballots for Audi<ng Cynthia Sturton, Eric Rescorla, David Wagner 1 Elec<on Audits are Important Source: Joe Hall 2 Audi<ng Methods Precinctbased: Standard prac<ce
SLIDE 2
SLIDE 3
Audi<ng Methods
- Precinct‐based:
– Standard prac<ce – Choose a sample of precincts to audit – Every ballot in a sampled precinct is audited
- Ballot‐based1,2,3:
– Newer idea – Choose a sample of ballots to audit – Sample from the set of all ballots, across precincts
- 1. Neff, C. A., Dec. 2003.
- 2. Johnson, K. C., Oct. 2004.
- 3. Calandrino, J. A., Halderman, J. A., and Felten, E. W., EVT 2007.
3
SLIDE 4
Ballot‐based vs. Precinct‐based
- Ballot‐based audi<ng is more efficient
– Confidence based on number of audit units rather than number of ballots
- E.g., Virginia 2006 elec<on results1
– Ballot‐based audi<ng would have required the recount of between 1/17 to 1/400 as many ballots as precinct‐based audi<ng did.
- Our focus is on ballot‐based audi<ng
4
- 1. Calandrino, J. A., Halderman, J. A., and Felten, E. W., EVT 2007.
SLIDE 5
Audit Sampled Ballots Verify Tabula<on
How ballot‐based audi<ng works
Returned Ballots Scanner Cast Vote Records (CVR) Elec<on Management So`ware Scanned Ballots Observer Tallies
SLIDE 6
A Challenge for Ballot‐based Audi<ng:
Finding the sampled ballot
- Key steps of ballot‐based audi<ng:
- 1. Picking cast vote record
- 2. Finding paper ballot
- 3. Compare paper ballot to cast vote record
- Requires a way to link each cast vote record to
its paper ballot
- Different proposals do this in different ways
6
SLIDE 7
Finding the Sampled Ballot Approach #1:
- Approach:
– Pre‐printed serial number
- Advantages:
– Conceptually simple
- Disadvantages:
– Violates privacy – Scanners require modifica<on ‐ so`ware – Finding par<cular ballot may be slow
7
SLIDE 8
Finding the Sampled Ballot Approach #2:
- Approach:
– Serial number stamped on a`er elec<on
- Advantages:
– Protects privacy – More efficient ballot selec<on
- Disadvantages:
– Scanners require modifica<on – so`ware & hardware – Modifies already‐voted ballots
8
SLIDE 9
Our Contribu<on
- Explicit serial number not necessary
- Loca<on in stack + Stack number =
Implicit serial number
9
SLIDE 10
Finding the Sampled Ballot Approach #3:
- Approach:
– Hand count to find implicit serial numbers
- Advantages:
– Protects privacy – No scanner modifica<on required – Voted ballots are not modified
- Disadvantages:
– Finding par<cular ballot may be slow – Possibility for human error
10
SLIDE 11
Finding the Sampled Ballot Approach #4:
- Approach:
– Use ballot weight to find implicit serial numbers
- Advantages:
– Protects privacy – No scanner modifica<on required – Voted ballots are not modified – Faster than hand coun<ng
- Disadvantages:
– Possibility for selec<on error
11
SLIDE 12
5 6 7 8 9 10 11 12 4 3 2 1
Ballot Stack Flipping Index into the stack by finding the sub‐stack with the correct number of ballots. Scale
12
SLIDE 13
A coun<ng scale efficiently counts the number of ballots in a stack
13
SLIDE 14
Selec<on Experiment
- Methodology
– 50kg x 0.002kg coun<ng scale – 350 Ballots – calibra<on and selec<on
- Results
– 20 Trials – Longest <me, 31 seconds (early trial) – All trials resulted in correct ballot selec<on
14
SLIDE 15
Sources of Selec<on Error
- Scale error
- Varia<on in ballot weights
- Mis‐es<ma<ng mean ballot weight
15
SLIDE 16
Projected Selec<on Error
- Calculate es<mated mean ballot weight
– 1000 ballots sampled with replacement
- Generate stacks of 500 ballots
- For each posi<on i in the stack, would we
correctly es<mate stack size?
- 100,000 trials
16
SLIDE 17
Simulated Error Rate Resul<ng from Varia<on in Ballot Mass
17
SLIDE 18
Limita<ons of this Research
- Unknown:
– Varia<on in weight of voted ballots – Homogeneity of ballot weight distribu<on across different boxes of ballots – Prac<cality of keeping ballot stack order – End‐to‐end efficiency of scheme
18
SLIDE 19
Conclusion
- We present a new scheme to enable ballot‐
based audi<ng
- Advantages over prior schemes
– Compa<ble with legacy hardware – No modifica<on of voted ballots
- A promising idea, more research warranted
19
SLIDE 20
End
20
SLIDE 21
Varia<on in Ballot Weight
21
Weight (grams) Box A Weight (grams) Box B Number of Ballots Number of Ballots
SLIDE 22
Ballot Weight Varia<on Accumulates
22
Total Weight Total Weight Probability Probability