1 of 20

CMSC839C / INST878D -

Governing Algorithms and Algorithmic Governance

Class #8:

Balancing Privacy and Data Use in the U.S. Census

​

2 of 20

Class Project: Checkpoint #2 due Oct 22

  • Checkpoint 2 (5% of class grade): Submit a detailed summary of your proposed policy intervention.

3 of 20

Main Points of Today

  • There are times when the needs for data are legitimate and the need for privacy are legitimate

​

  • Computing techniques have something significant to offer when these tensions arise

​

  • If deployments of these computing techniques don’t take sociotechnical dynamics into account, they can be disasters

​

  • DP and the census might be the most successful CS-driven policy intervention in US history

4 of 20

What is privacy?

5 of 20

Government use of statistics

Article 1 Section 2 US Constitution:

https://censusreporter.org/profiles/14000US24033980000-census-tract-9800-prince-georges-md/

6 of 20

In CI language, what should we expect from this data collection?

7 of 20

Database Reconstruction Theorem

​

​

​

Theorem (informally) [Dinur and Nissim 2003]: Given enough*, sufficiently accurate* queries to a dataset, you can reconstruction the entire dataset

8 of 20

Differential Privacy

Local DP Experience - “Is your phone an iPhone?”

​

  • Roll Your Die twice
    • If ODD: tell me the truth
    • If EVEN: tell me the result from the 2nd die roll
      • ODD: Yes
      • Even: No

9 of 20

Differential Privacy

10 of 20

Differential Privacy

Central DP

​

​

​

11 of 20

Question: is DP the “right” privacy notion for Census?

Why?

12 of 20

What happened next?

13 of 20

The Negative: Ruggles and Van Ripper on Baselines

14 of 20

The Positive: Cohen, Moon, and Suwal on Comparisons

15 of 20

Parsing the Debate: boyd and Sarathy

16 of 20

Looking to 2030: What should we do next?

  • Process: are there better engagement strategies?

​

  • Technology: are there better definitions we care use?

17 of 20

Take Aways:

  • If we think of this story as a success story, what should we take away for other interventions?

​

  • If we think of this story as a failure story, what should we take away for other interventions?

18 of 20

Gabe’s opinion:

Without better tools to enable engagement around epsilon, DP cannot function

19 of 20

Gabe’s opinion:

20 of 20

Next Class: PETs more generally