Data Ethics: Algorithmic Privacy and Scarce Resource Allocation

CDTO Campus 2024: Shaping the Digital Future

Jeff Jacobs

Georgetown University
jj1088@georgetown.edu

Thursday, 8 August 2024

Hello!

My Goal

I want to provide some tools from data ethics, for your toolbox
If one of the tools is not relevant, I hope another will be!
If one of the tools is very relevant, I can provide more resources (books, articles) for you!

Who Am I? How Can I Help You?

💻		🌏
MS in Computer Science, Stanford University	→	PhD in International Relations, Columbia University
Assistant Teaching Professor, Data Science and Analytics	→	Courtesy Teaching Professor, McCourt School of Public Policy

Cross-listed course: DSAN 5450: Data Ethics and Policy
- All materials online + open-source, available at that link!

Dissertation: “Our Word is Our Weapon”

Cold War arms shipments (SIPRI) vs. propaganda (Печать СССР): here, to 🇪🇹

I only know the most basic words 🙈 but I will try my best!

Data Ethics in Times of War

🇵🇸	🇺🇦
2015-2023: Taught Mobile App Development in West Bank+Gaza	So when I learned of дія, I thought some lessons/skills could apply!

Ethical Frameworks: Implicit and Explicit

We are always using ethical frameworks, even when we don’t realize it!

Cake Cutting

You and a friend are very hungry, and you come across a tasty cake!
How do you divide it?

Ethical Frameworks

Your instinct may be: divide equally, so that you and your friend are treated equally
But… are you treated equally?
What if your friend is 10 times hungrier?
What if your stomach is sensitive to sugar, so you can’t eat more than \(1/4\)?

Math to the Rescue… For Now!

For \(N = 2\), simple mechanism for envy-free division:

Анастасія cuts the cake into 2 pieces however she wants
Вероніка chooses which piece to eat
…Давай розпочнемо цю вечірку!¹

Ethical Complexity: 3 People

Ethical Complexity: \(n\) People

We consider the well-studied cake cutting problem in which the goal is to find an envy-free allocation based on queries from \(n\) agents. The problem has received attention in computer science, mathematics, and economics. It has been a major open problem whether there exists a discrete bounded envy-free protocol. We resolve the problem with a discrete bounded envy-free protocol for any number of agents. The maximum number of queries required by the protocol is \(n^{n^{n^{n^{n^n}}}}\). (Aziz and Mackenzie 2016)

Ethical Framework Implicit in Learning “Who You Are”!

Code

import spacy
nlp = {
  'en': spacy.load("en_core_web_sm"),
  'uk': spacy.load("uk_core_news_sm")
}
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from wordcloud import WordCloud
portfolio_df = pd.read_csv("Student_portfolios.csv")
portfolio_df = portfolio_df[~pd.isna(portfolio_df['Bio'])].copy()
bio_list_en = list(portfolio_df['Bio'].values)
# Preprocess bios individually
def clean_bio(raw_bio_str, lang):
  lang_nlp = nlp[lang]
  bio_doc = lang_nlp(raw_bio_str)
  bio_tokens = [t.text for t in bio_doc if not t.is_stop]
  bio_sent = " ".join(bio_tokens)
  return bio_sent
def clean_bios(raw_bio_list, lang):
  cleaned_sents = []
  for cur_bio in raw_bio_list:
    bio_sent = clean_bio(cur_bio, lang)
    cleaned_sents.append(bio_sent)
  return cleaned_sents
cleaned_bio_list_en = clean_bios(bio_list_en, 'en')
def combine_bios(bio_list):
  bio_str = "\n".join(bio_list)
  return bio_str
bio_str_en = combine_bios(cleaned_bio_list_en)
wc_options = {
  'background_color': 'white',
  'font_path': 'assets/Roboto.ttf',
  'scale': 8,
  'random_state': 2024,
  'width': 500,
  'height': 400,
}
def display_wordcloud(text_str):
  cloud = WordCloud(**wc_options)
  cloud.generate(text_str)
  plt.imshow(cloud, interpolation='bilinear')
  plt.axis("off")
display_wordcloud(bio_str_en)

Code

bio_list_uk = list(portfolio_df['Bio_uk'].values)
cleaned_bio_list_uk = clean_bios(bio_list_uk, 'uk')
bio_str_uk = combine_bios(cleaned_bio_list_uk)
display_wordcloud(bio_str_uk)

How might this be biased, if we want equal contribution?

Now With Fake Jeff Bio

Code

jj_sent_en = "Jeff teaches Data Ethics at Georgetown University."
jj_bio_en = " ".join([jj_sent_en] * 100)
print(jj_bio_en[:230] + "...")

Jeff teaches Data Ethics at Georgetown University. Jeff teaches Data Ethics at Georgetown University. Jeff teaches Data Ethics at Georgetown University. Jeff teaches Data Ethics at Georgetown University. Jeff teaches Data Ethics a...

Code

bio_list_fake_en = cleaned_bio_list_en.copy()
cleaned_jj_bio_en = clean_bio(jj_bio_en, 'en')
bio_list_fake_en.append(cleaned_jj_bio_en)
bio_str_fake_en = combine_bios(bio_list_fake_en)
bio_cloud_fake_en = WordCloud(**wc_options)
bio_cloud_fake_en.generate(bio_str_fake_en)
plt.imshow(bio_cloud_fake_en, interpolation='bilinear')
plt.axis("off")

Code

jj_sent_uk = "Джеф викладає етику даних у Джорджтаунському університеті."
jj_bio_uk = " ".join([jj_sent_uk] * 100)
print(jj_bio_uk[:200] + "...")

Джеф викладає етику даних у Джорджтаунському університеті. Джеф викладає етику даних у Джорджтаунському університеті. Джеф викладає етику даних у Джорджтаунському університеті. Джеф викладає етику дан...

Code

bio_list_fake_uk = cleaned_bio_list_uk.copy()
cleaned_jj_bio_uk = clean_bio(jj_bio_uk, 'uk')
bio_list_fake_uk.append(cleaned_jj_bio_uk)
bio_str_fake_uk = combine_bios(bio_list_fake_uk)
bio_cloud_fake_uk = WordCloud(**wc_options)
bio_cloud_fake_uk.generate(bio_str_fake_uk)
plt.imshow(bio_cloud_fake_uk, interpolation='bilinear')
plt.axis("off")

Normalizing Counts

This is where ethical framework comes into play, whether implicit or explicit!
Normalized: each person \(i\) gets weight \(\omega_i = 1/N\)

Code

# Each bio's freqs should count 1/N in overall
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.preprocessing import normalize
cv_en = CountVectorizer(
  min_df=1,
  # stop_words='english'
)
dtm_en_sparse = cv_en.fit_transform(bio_list_fake_en)
dtm_en = dtm_en_sparse.todense()
#print(dtm.shape)
row_sums_en = dtm_en.sum(axis=1)
#print(row_sums.shape)
dtm_r0_en = dtm_en[0,:]
r0_sum_en = dtm_r0_en.sum()
r0_norm_en = dtm_r0_en / r0_sum_en
# Now use `normalize()`
dtm_norm_en = normalize(np.asarray(dtm_en), axis=1, norm='l1')
# And compute col sums of normalized DTM
dtm_colsums_en = dtm_norm_en.sum(axis=0)
# And combine with the words in a df
freq_df_en = pd.DataFrame({
  'term': cv_en.get_feature_names_out(),
  'freq': dtm_colsums_en
})
#freq_df['term'].values

Code

freq_dict_en = freq_df_en.set_index('term').to_dict(orient='dict')['freq']
bio_cloud_eq_en = WordCloud(**wc_options)
bio_cloud_eq_en.generate_from_frequencies(freq_dict_en)
plt.imshow(bio_cloud_eq_en, interpolation='bilinear')
plt.axis("off")

Code

cv_uk = CountVectorizer(
  min_df=1,
)
dtm_uk_sparse = cv_uk.fit_transform(bio_list_fake_uk)
dtm_uk = dtm_uk_sparse.todense()
#print(dtm.shape)
row_sums_uk = dtm_uk.sum(axis=1)
#print(row_sums.shape)
dtm_r0_uk = dtm_uk[0,:]
r0_sum_uk = dtm_r0_uk.sum()
r0_norm_uk = dtm_r0_uk / r0_sum_uk
# Now use `normalize()`
dtm_norm_uk = normalize(np.asarray(dtm_uk), axis=1, norm='l1')
# And compute col sums of normalized DTM
dtm_colsums_uk = dtm_norm_uk.sum(axis=0)
# And combine with the words in a df
freq_df_uk = pd.DataFrame({
  'term': cv_uk.get_feature_names_out(),
  'freq': dtm_colsums_uk
})
#freq_df['term'].values

Code

freq_dict_uk = freq_df_uk.set_index('term').to_dict(orient='dict')['freq']
bio_cloud_eq_uk = WordCloud(**wc_options)
bio_cloud_eq_uk.generate_from_frequencies(freq_dict_uk)
plt.imshow(bio_cloud_eq_uk, interpolation='bilinear')
plt.axis("off")

Equality of What?

Is \(\omega_i = 1/N\) the “correct” choice?
Should we ensure each person/ministry gets \(\geq 1\) big word
Maybe it should be 50% public sector, 50% private sector
Higher \(\omega\) for those sitting closer to me: \(\omega_i = \frac{1}{\text{dist}(\text{Jeff},i)}\)
…How to decide? Objective Function \(f\) + Constraints \(g\)!

In General:

\[ \begin{align*} \max_{c \mkern1.0mu \in \mkern1.0mu \text{Choices}} \; & f(c) = \text{Goodness of }c \\ \text{subject to } & g(c) = \text{Constraints on }c \end{align*} \]

Our Example:

\[ \begin{align*} \max_{\omega \mkern1.0mu \in \mkern1.0mu [0,1]^N} \; & f(\omega) = \textstyle\sum_{i=0}^{N}\omega_i u_i \\ \text{subject to } &\textstyle\sum_{i=0}^{N}\omega_i = 1 \end{align*} \]

Resource Allocation

Background: Indochina Mine Defusal

Using GIS to allocate mine defusal teams to regions of Indochina mined by the US
What information do we need about each region?
- Absolute population? Population density? Mines per person?
- Can equipment be transported there?

Incremental Model

As a Minimum Viable Product (MVP), first just consider the non-constrained optimization
Add in constraints incrementally—as you add, evaluate feasibility relative to goals (for example, how close can you get to MVP outcome?)
Iterative process \(\implies\) visualization can be crucially helpful!

Interactive Examples!

ObservableJS Interactive Dashboard

Problem 1: Divisibility

Can we split our resource infinitely?
\(1/2\) of a gas mask doesn’t help us! ☠️

Problem 2: Distance Metrics

Euclidean / “Straight line” distance not always the “right” measure!
Allocation Challenge 2: Distance-Based Decay

Problem 3: Capabilities

How effectively can people convert resources \(\leadsto\) utility?
Digital: internet availability
Physical: nutrition during pregnancy
Johns Hopkins Medicine: 300 more calories per day

Privacy-Preserving Computation

Disclosure Avoidance

1990, 2000, and 2010 Census: “Data Swapping”:

We inject “noise” into the data by swapping records for certain households with those from households with similar characteristics in a nearby area. The Census does not release information about its specific methods for swapping. While this confidentiality around swapping techniques is important to protect against disclosure, it means that the practice is not transparent to data users

The Death of Data Swapping

But… Штучний інтелект, Великі дані, etc… 🤖

If traditional disclosure avoidance techniques were applied to the 2020 Census data, the amount of noise required to protect against new attacks would make census data unfit for most uses

\(\implies\) We need a framework for quantifying “acceptable” privacy loss (tradeoff!)

Privacy Tradeoff

Figure 1.3: “The Accuracy-Privacy Loss Tradeoff”, from US Census Bureau (2021), Disclosure Avoidance for the 2020 Census: An Introduction

Differential Privacy

Implemented for the 2020 US Census
Key idea: Choose \(\varepsilon_i\) based on how much you trust person \(i\), who wants to see your data.

Figure 8.1: “The Privacy-Loss Budget (Epsilon) Acts as a Dial on the Level of Noise”, from US Census Bureau (2021), Disclosure Avoidance for the 2020 Census: An Introduction

Why Can’t We Just Restrict Queries to Aggregated Statistics?

The answer: Differencing Attacks 😰

Query 1: Number of народних депутатів in Верховна Рада with disease \(X\)
Query 2: Number of народних депутатів in Верховна Рада, except Голова, with disease \(X\)

Real Data: селище Понорниця

Example with Real Data

громада	True Population			Noise			Noisy Counts			Post-Processed
громада	<18	18+	Total	<18	18+	Total	<18	18+	Total	<18	18+	Total
Зелена Поляна	20	57	77	0	-4	+2	20	53	(≠)79	23 (+3)	54 (+1)	77 (-2)
Рихли	30	82	112	-3	+2	+3	27	84	115	28 (+1)	85 (+1)	113 (-2)
Великий Ліс	20	86	106	-2	+1	+1	18	87	107	17 (-1)	87	104 (-3)
Розльоти	150	155	305	0	+2	0	150	157	305	150	156 (-1)	306 (+1)
Total	220	380	600	-5	+1	+6	215	381	606	218	382	600

The Magic (Noise!)

If noise values \(\nu\) drawn i.i.d. from Laplace distribution (pdf):

\[ M(D) \triangleq \text{Lap}(x; \varepsilon) = \frac{1}{2\varepsilon}\exp\mkern-2.5mu\left[-\frac{|x|}{\varepsilon}\right] \]

Can mathematically guarantee \(\varepsilon\)-differential privacy!¹

\[ \Pr\left[M(D) \in S\right] \leq e^\varepsilon \times \Pr\mkern-2.5mu\left[M(D') \in S\right] \]

Can quantify privacy loss \(\mathcal{L}(\gamma)\) from observing \(\gamma\):

\[ \mathcal{L}(\gamma) = \ln\left[\frac{\Pr[M(D) = \gamma]}{\Pr[M (D') = \gamma]}\right] \]

The Metaphor

Seurat, A Sunday Afternoon on the Island of La Grande Jatte, Wikimedia Commons

References / Resources

Aziz, Haris, and Simon Mackenzie. 2016. “A Discrete and Bounded Envy-Free Cake Cutting Protocol for Any Number of Agents.” IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS), 416–27. https://doi.org/10.1109/FOCS.2016.52.

Chiappero-Martinetti, Enrica, Siddiqur Osmani, and Mozaffar Qizilbash. 2020. The Cambridge Handbook of the Capability Approach. Cambridge University Press.

Dwork, Cynthia. 2014. “Differential Privacy: A Cryptographic Approach to Private Data Analysis.” In Privacy, Big Data, and the Public Good: Frameworks for Engagement, edited by Helen Nissenbaum, Julia Lane, Stefan Bender, and Victoria Stodden. Cambridge University Press. https://doi.org/10.1017/CBO9781107590205.018.

Lane, Julia, Victoria Stodden, Stefan Bender, and Helen Nissenbaum. 2014. Privacy, Big Data, and the Public Good. Cambridge University Press.

Robert, Amélie. 2016. “At the Heart of the Vietnam War: Herbicides, Napalm and Bulldozers Against the A Lưới Mountains.” Journal of Alpine Research Revue de géographie Alpine, nos. 104-1 (April). https://doi.org/10.4000/rga.3266.

Sen, Amartya. 1985. Commodities and Capabilities. OUP India.

Shahid, Rizwan, Stefania Bertazzon, Merril L. Knudtson, and William A. Ghali. 2009. “Comparison of Distance Measures in Spatial Analytical Modeling for Health Service Planning.” BMC Health Services Research 9 (1): 200. https://doi.org/10.1186/1472-6963-9-200.

Q & A

Thank You!

Appendix Slides

Sensitivity

Consider a real-valued function \(f\). The (worst-case, or global) sensitivity of \(f\) is the maximum absolute value by which the addition or deletion of a single database row can change the value of \(f\):

\[ \Delta f = \max_{D, D'}|f(D) - f(D')| \]

Queries of the form “How many people in the database are over six feet tall?” have sensitivity \(\Delta f = 1\), since the presence or absence of any individual in \(D\) can affect the true answer by at most 1. Thus, the Laplace mechanism returns the true count perturbed by a random draw from \(\text{Lap}(x; 1/\varepsilon)\). (Dwork 2014)

Data Ethics: Algorithmic Privacy and Scarce Resource Allocation

Hello!

My Goal

Who Am I? How Can I Help You?

Dissertation: “Our Word is Our Weapon”

Data Ethics in Times of War

Ethical Frameworks: Implicit and Explicit

Cake Cutting

Ethical Frameworks

Math to the Rescue… For Now!

Ethical Complexity: 3 People

Ethical Complexity: \(n\) People

Ethical Framework Implicit in Learning “Who You Are”!

Now With Fake Jeff Bio

Normalizing Counts

Equality of What?

Resource Allocation

Background: Indochina Mine Defusal

Incremental Model

Sidebar: Україна in ArcGIS

Interactive Examples!

Problem 1: Divisibility

Problem 2: Distance Metrics

Problem 3: Capabilities

Privacy-Preserving Computation

Disclosure Avoidance

The Death of Data Swapping

Privacy Tradeoff

Differential Privacy

Why Can’t We Just Restrict Queries to Aggregated Statistics?

Real Data: селище Понорниця

Example with Real Data

The Magic (Noise!)

The Metaphor

References / Resources

Q & A

Thank You!

Appendix Slides

Sensitivity