CSE 163

Fairness & Privacy�

Suh Young Choi��🎶 Listening to: Death’s Door Soundtrack�💬 Before Class: What has been your favorite Before-Class question so far?�

Announcements

Checkpoint 3 and Learning Reflection 4 due tonight

​

Resubmission Cycle closes tomorrow (late submissions for HW 5 are allowed here)

​

Course Evaluations opening on Wednesday

​

Project Report and Code due Thursday

​

​

​

​

​

2

Checking in

Resubmissions

• Two more cycles guaranteed
• 8/6 – 8/12 (closes tomorrow)
• 8/13 – 8/19
• If 60% of the class fills out the course eval by 8/19, then there will be a bonus resubmission period from 8/20 to 8/22
• You can resubmit any ONE take-home assessment!

​

Projects and Pacing

• Course staff can help you figure out what is reasonable to complete
• It’s OK if your work doesn’t meet your expectations—show us how far you got!

​

3

This Time

• Defining Fairness
• Quantifying error rates
• Privacy

Last Time

• Machine learning with images
• Ethics of ML

4

Group Fairness

Intent: Avoid discrimination against a particular group, as to avoid membership in the group negatively impact outcomes for people in that group.

• Does not say which groups to protect, that’s a decision of policy and social norms.
• Can be extended to notions of belonging to multiple identities (e.g., intersectionality), but we focus on protecting a single property at this time

​

Usually defined in terms of mistakes the system might make

5

Definitions of Fairness

Equality of False Negatives (equal opportunity): False negative rate should be similar across groups

​

​

​

​

​

​

​

​

​

​

6

* Many others exist, many are in the form of equations on this confusion matrix! There are other notions of fairness too!

College admission example: P = Successful in college, N = Not successful in college

Definitions of Fairness

Equality of False Positives (predictive equality): False positive rate should be similar across groups

7

* Many others exist, many are in the form of equations on this confusion matrix! There are other notions of fairness too!

College admission example: P = Successful in college, N = Not successful in college

Human Choice

There is no one “right” definition for fairness. They are all valid and are simply statements of what you believe fairness means in your system.

​

It’s possible for definitions of fairness to contradict each other, so it’s important that you pick the one that reflects your values.

​

Emphasizes the role of people in the process of fixing bias in ML algorithms.

8

Tradeoff Between Fairness and Accuracy

We can’t get fairness for free, generally finding a more fair model will yield to one that is less accurate.

• Intuition: We saw lots of examples where bias was a byproduct of an “accurate” model since that model was not trained with fairness in mind.

​

Can quantify this tradeoff with Pareto Frontiers

9

Pareto Frontiers

10

Fairness Worldviews

Example: College admissions

We want to measure abstract qualities about a person (e.g., intelligence or grit), but real life measurements may or may not measure abstract qualities well.

Only have access to Observed Space and we hope it’s a good representation of the Construct Space.

11

Worldview 1: WYSIWYG

Worldview 1: What You See is What You Get (WYSIWYG)

• Assumes the Observed Space is a good representation of the Construct Space.

​

Under this worldview, can guarantee individual fairness. Individual fairness says if two people are close in the Construct Space, they should receive similar outcomes.

• Easy to verify under WYSIWYG since you can use the Observed Space as a good representation of the Construct Space.

12

Worldview 2: Structural Bias + WAE

Worldview 2: Structural Bias and We’re All Equal (WAE)

• Assumes systematic or social systems make different groups that look similar in the Construct Space look more different in the Observed Space.
• Example: SAT Scores for one group might be artificially high due to better ability to afford SAT prep. Factors outside of qualities of interest now affect our measurements. So we assume any observed differences between groups are systematic factors, rather than inherent factors since WAE.

​

Goal in this worldview is to ensure non-discrimination so that someone isn’t negatively impacted by simply being a member of a particular group.

• This is the implicit assumption we we were making when discussing notions of group fairness earlier

13

Contrasting Worldviews

Unfortunately there is no way to tell which worldview is right for a given problem (no access to Construct Space). The worldview is a statement of beliefs.

​

WYSIWYG can promise individual fairness but methods of non-discrimination will be individually unfair under this worldview.

​

Structural Bias + WAE can promise non-discrimination. Methods of individual fairness will lead to discrimination (since using biased data as our proxy for closeness will lead to a skewed notion of individually fair).

14

Anonymous Data Isn’t

Mid 1990s, insurance group in Massachusetts published anonymous records of hospital visits with attributes like name, address, social security removed but left in demographic information.

​

Turns out this data release was not so anonymous!

• Latanya Sweeney was able to link demographic information in hospital data with voter rolls. Found which hospital record corresponded to governor.

​

Sweeney estimates 87% of the US is uniquely identified by knowing 1) date of birth, 2) sex, and 3) zip code.

15

k-anonymity

K-anonymity: A first definition of privacy by Sweeney that requires every query results in at least k people in the dataset.

• Achieved by removing columns or fuzzing values

​

​

​

​

​

​

​

​

​

Weakness: Fails under composition

​

​

​

16

Differential Privacy

A stronger notion of privacy that guarantees how much information you can learn about a person.

​

Consider two worlds, one where A participates in a study and one where they don’t. If results of the study are similar, we say it respects differential privacy.

​

​

17

Differential Privacy

Say an algorithm or analysis is 𝜀-differentially private if results with or without any single person in the dataset are “at most 𝜀” off.

• Defining how close results are is a little complex, but is a statement of probabilities
• If 𝜀 = 0, require results to be exactly the same
• If 𝜀 is small, require results to be very similar
• If 𝜀 is large, require more deviation in results (less privacy)

​

Two methods for commonly achieving 𝜀-differential privacy

• Jittering Result (Laplace Mechanism)
• Randomized Response

18

Jittering

Take result of analysis and add a small amount of random noise to result.

• Example: Report average age of census but add a small random number to it

​

Specifically if you add noise that follows a Laplace distribution with parameter 𝜀, you can achieve 𝜀-differential privacy.

• See below for 𝜀=0.5 (red dashes), 𝜀=1 (blue solid), 𝜀=2 (purple dots)

19

Randomized Response

What if we don’t trust the data collector with our data?

• Even with differentially private statistics published, they still have access to the raw data. What if they get hacked?

�Change the differential privacy mechanism to be done locally rather than centrally!

​

Differentially Private Polling Procedure:

• Call up person. Ask them to flip a coin (don’t tell us result)
• If Heads, tell us their honest answer to question (“Yes”/”No”)
• If Tails, flip coin again
• If Heads, report “Yes”
• If Tails, report “No”

�Key idea: Can learn aggregate trends without knowing true result of individual

20

Randomized Response Analysis

Key property: People tell the truth ¾ of the time and lie ¼ of the time. ½ of the time they are honest, and then half of the time they tell us a random answer that lines up with the truth.

​

To see why this work, suppose we know the answer is “Yes” for ⅓ of people. How many “Yes” responses would we expect in this procedure?

• ⅓ of the population has true answer “Yes”. ¾ of them will tell us the truth so we will get a total of ¼ of responses being honest “Yes”es
• ⅔ of the population has the true answer “No”, but ¼ of the time they will randomly tell us “Yes”. This means we would expect ⅙ of the population to lie and tell us “Yes”
• Total of “Yes” received (on average): ¼ + ⅙ = 5/12

​

In general, work backwards to solve for underlying probability

​

​

21

Before Next Time

• Read through Lesson 22
• CP3 + LR 6 due tonight
• Resub cycle closes tomorrow (last day to resubmit HW3)

​

Next Time

• Statistical Testing
• Research Methods

​

22