SlideShare a Scribd company logo
1 of 199
Download to read offline
Privacy-preserving
Data Mining in Industry
WWW 2019 Tutorial
May 2019
Krishnaram Kenthapadi (AI @ LinkedIn)
Ilya Mironov (Google AI)
Abhradeep Thakurta (UC Santa Cruz)
https://sites.google.com/view/www19-privacy-tutorial
Fairness Privacy
Transparency Explainability
Fairness Privacy
Transparency Explainability
Related WWW’19 sessions:
1.Tutorial: Designing Equitable Algorithms for the Web
2.Tutorial: Economic Theories of Distributive Justice for Fair Machine Learning
3.Tutorial: Socially Responsible NLP
4.Tutorial: Fairness-aware Machine Learning in Practice
5.Tutorial: Explainable Recommendation and Search
6.Workshop: FATE and Society on the Web
7.Session: Fairness, Credibility, and Search (Wednesday, 10:30 – 12:30)
8.Session: Privacy and Trust (Wednesday, 16:00 – 17:30)
9.Special Track: Designing an Ethical Web (Friday)
“Privacy by Design” for AI products
Outline / Learning Outcomes
• Privacy breaches and lessons learned
• Evolution of privacy techniques
• Differential privacy: definition and techniques
• Privacy techniques in practice: Challenges and Lessons Learned
• Google’s RAPPOR
• Apple’s differential privacy deployment for iOS
• Privacy in AI @ LinkedIn (Analytics framework & LinkedIn Salary)
• Key Takeaways
Privacy: A Historical Perspective
Evolution of Privacy Techniques and Privacy Breaches
Privacy Breaches and Lessons Learned
Attacks on privacy
•Governor of Massachusetts
•AOL
•Netflix
•Web browsing data
•Facebook
•Amazon
•Genetic data
born July 31, 1945
resident of 02138
Massachusetts Group Insurance Commission (1997):
Anonymized medical history of state employees (all
hospital visits, diagnosis, prescriptions)
Latanya Sweeney (MIT grad student): $20 – Cambridge
voter roll
William Weld vs Latanya Sweeney
64
%uniquely identifiable with
ZIP + birth date + gender
(in the US population)
Golle, “Revisiting the Uniqueness of Simple Demographics in the US Population”,
Attacker's Advantage
Auxiliary information
August 4, 2006: AOL
Research publishes
anonymized search logs
of 650,000 users
August 9:
New York Times
AOL Data Release
Attacker's Advantage
Auxiliary information
Enough to succeed on a small fraction of inputs
Netflix Prize
Oct 2006: Netflix announces Netflix
Prize
• 10% of their users
• average 200 ratings per user
Narayanan, Shmatikov (2006):
Netflix Prize
Deanonymizing Netflix Data
Narayanan, Shmatikov, Robust De-
anonymization of Large Datasets (How to
Break Anonymity of the Netflix Prize
Dataset), 2008
● Noam Chomsky in Our Times
● Farenheit 9/11
● Jesus of Nazareth
● Queer as Folk
Key idea:
● Similar intuition as the attack on medical records
● Medical records: Each person can be identified
based on a combination of a few attributes
● Web browsing history: Browsing history is unique for
each person
● Each person has a distinctive social network  links
appearing in one’s feed is unique
● Users likely to visit links in their feed with higher
probability than a random user
● “Browsing histories contain tell-tale marks of identity”
Su et al, De-anonymizing Web Browsing Data with Social Networks, 2017
De-anonymizing Web Browsing Data with Social Networks
Attacker's Advantage
Auxiliary information
Enough to succeed on a small fraction of inputs
High dimensionality
Ad targeting:
Korolova, “Privacy Violations Using Microtargeted Ads: A Case Study”, PADM
Privacy Attacks On Ad Targeting
10 campaigns targeting 1 person (zip code, gender,
workplace, alma mater)
Korolova, “Privacy Violations Using Microtargeted Ads: A Case Study”, PADM
Facebook vs Korolova
Age
21
22
23
…
30
Ad Impressions in a week
0
0
8
…
0
10 campaigns targeting 1 person (zip code, gender,
workplace, alma mater)
Korolova, “Privacy Violations Using Microtargeted Ads: A Case Study”, PADM
Facebook vs Korolova
Interest
A
B
C
…
Z
Ad Impressions in a week
0
0
8
…
0
● Context: Microtargeted Ads
● Takeaway: Attackers can instrument ad campaigns to
identify individual users.
● Two types of attacks:
○ Inference from Impressions
○ Inference from Clicks
Facebook vs Korolova: Recap
Attacker's Advantage
Auxiliary information
Enough to succeed on a small fraction of inputs
High dimensionality
Active
Items frequently bought together
Bought: A B C D E
Z: Customers Who Bought This Item Also Bought
Calandrino, Kilzer, Narayanan, Felten, Shmatikov, “You Might Also Like: Privacy Risks of Collaborative
Attacking Amazon.com
A C D E
Attacker's Advantage
Auxiliary information
Enough to succeed on a small fraction of inputs
High dimensionality
Active
Observant
Homer et al., “Resolving individuals contributing trace
amounts of DNA to highly complex mixtures using high-
density SNP genotyping microarrays”, PLoS Genetics,
2008
Genetic data
Reference population
Bayesian Analysis
“In all mixtures, the identification
of the presence of a person’s
genomic DNA was possible.”
Zerhouni, NIH Director:
“As a result, the NIH has removed from
open-access databases the aggregate
results (including P values and genotype
counts) for all the GWAS that had been
available on NIH sites”
… one week later
Attacker's Advantage
Auxiliary information
Enough to succeed on a small fraction of inputs
High dimensionality
Active
Observant
Clever
Negative Results
Dinur-Nissim
0 1 1 0 1 0 0 0 1 1 0 1Data
query: 𝚺
Dinur-Nissim 2003:
If error is o(√n), then reconstruction is possible up to n−o(n)
...even if 23.9% of errors are arbitrary [DMT07]
...even with O(n) queries [DY08]
Dwork-Naor
Tore Dalenius desideratum (aka as “semantic security”):
“Access to a statistical database should not enable one to
learn anything about an individual that could not be learned
without access.” (1977)
Dwork-Naor (~2006):
If the database teaches us anything, there is always some auxiliary
information that breaks Dalenius desideratum.
Differential Privacy
Curator
Defining Privacy
Curator
Defining Privacy: Fool's Errand
Defining Privacy
38
CuratorCurator
+ your data
- your data
Differential Privacy
39
Databases D and D′ are neighbors if they differ in one person’s data.
Differential Privacy: The distribution of the curator’s output M(D) on database
D is (nearly) the same as M(D′).
CuratorCurator
+ your data
- your data
Dwork, McSherry, Nissim, Smith [TCC 2006]
ε-Differential Privacy: The distribution of the curator’s output M(D) on
database D is (nearly) the same as M(D′).
Differential Privacy
40
CuratorCurator
Parameter ε quantifies
information leakage
∀S: Pr[M(D)∊S] ≤ exp(ε) ∙ Pr[M(D′)∊S].
+ your data
- your data
Dwork, McSherry, Nissim, Smith [TCC 2006]
ε-Differential Privacy: The distribution of the curator’s output M(D) on
database D is (nearly) the same as M(D′).
Differential Privacy
41
CuratorCurator
Parameter ε quantifies
information leakage
∀S: Pr[M(D)∊S] ≤ exp(ε) ∙ Pr[M(D′)∊S]+𝛿.
Parameter 𝛿 gives
some slack
Dwork, Kenthapadi, McSherry, Mironov, Naor [EUROCRYPT 2006]
+ your data
- your data
42
f(D) f(D′)
— bad outcomes
— probability with record x
— probability without record x
“Bad Outcomes” Interpretation
● Prior on databases p
● Observed output O
● Does the database contain record x?
43
Bayesian Interpretation
Differential Privacy
● Robustness to auxiliary data
● Post-processing:
If M(D) is differentially private, so is f(M(D)).
● Composability:
Run two ε-DP mechanisms. Full interaction is 2ε-DP.
● Group privacy:
Graceful degradation in the presence of correlated inputs.
44
Differential Privacy: Laplace Mechanism
Define ℓ1-sensitivity of f: D→ℝn:
maxD,D′ ||f(D) − f(D′)||1 < 1,
then the Laplace mechanism
f(D) + Laplacen(1/ε)
offers ε-differential privacy.
Dwork, McSherry, Nissim, Smith, “Calibrating Noise to Sensitivity in Private Data Analysis”, TCC 2006
Differential Privacy: Gaussian Mechanism
If ℓ2-sensitivity of f:D→ℝn:
maxD,D′ ||f(D) − f(D′)||2 < 1,
then the Gaussian mechanism
f(D) + Nn(0, σ2)
offers (ε, δ)-differential privacy, where δ = ⅘·exp(-(εσ)2/2).
Dwork, Kenthapadi, McSherry, Mironov, Naor, “Our Data, Ourselves”, Eurocrypt 2006
What Differential Privacy Isn’t
● Algorithm, architecture, or a rule book
● Secure Computation: what not how
● All-encompassing guarantee: trends may be
sensitive too
Strava Fitness App
BBC: “Fitness app Strava lights up staff at military bases”
Differential Privacy: Takeaway points
• Privacy as a notion of stability of randomized algorithms in
respect to small perturbations in their input
• Worst-case definition
• Robust (to auxiliary data, correlated inputs)
• Composable
• Quantifiable
• Concept of a privacy budget
• Noise injection
Case Studies
Google’s RAPPOR
...Mountain View, 2014
Central Model
Curator
Local Model
Differential Privacy
ε-Differential Privacy: The distribution of the output M(D) on database
D is (nearly) the same as M(D′) for all adjacent databases D and D′:
∀S: Pr[M(D)∊S] ≤ exp(ε) ∙ Pr[M(D′)∊S].
Local Differential Privacy
ε-Differential Privacy: The distribution of the output M(D) on database
D is (nearly) the same as M(D′) for all adjacent databases D and D′:
∀S: Pr[M(D)∊S] ≤ exp(ε) ∙ Pr[M(D′)∊S].
Local-Differentially Private Mechanisms
● Stanley L. Warner, "Randomized response: a survey technique for
eliminating evasive answer bias", Journal of American Statistical
Association, March 1965.
● Arijit Chaudhuri, Rahul Mukerjee. Randomized
Response. Theory and Techniques. 1988.
Randomized Response (Warner 1965)
Q1: Are you a citizen of the United States?
Q2: Are you not a citizen of the United States?
𝜃 - the true fraction of citizens in the sample
Answer Q1 Answer Q2
p 1 − p
-DP
RAPPOR
Erlingsson, Pihur, Korolova. "RAPPOR: Randomized Aggregatable Privacy-Preserving
Ordinal Response." ACM CCS 2014.
RAPPOR: two-level randomized response
Can we do repeated surveys of sensitive attributes?
— Average of randomized responses will reveal a user’s true answer :-(
Solution: Memoize! Re-use the same random answer
— Memoization can hurt privacy too! Long, random bit sequence can
be a unique tracking ID :-(
Solution: Use 2-levels! Randomize the memoized response
RAPPOR: two-level randomized response
● Store client value v into bloom filter B using hash functions
● Memoize a Permanent Randomized Response (PRR) B′
● Report an Instantaneous Randomized Response (IRR) S
RAPPOR: two-level randomized response
● Store client value v into bloom filter B using hash functions
● Memoize a Permanent Randomized Response (PRR) B′
● Report an Instantaneous Randomized Response (IRR) S
f = ½
q = ¾ , p = ½
RAPPOR: Life of a report
Value
Bloom
Filter
PRR
IRR
“www.google.com”
Value
Bloom
Filter
PRR
IRR
“www.google.com”
P(1) =
0.25
P(1) =
0.75
RAPPOR: Life of a report
Value
Bloom
Filter
PRR
IRR
“www.google.com”
P(1) =
0.50
P(1) =
0.75
RAPPOR: Life of a report
Differential privacy of RAPPOR
● Permanent Randomized Response satisfies differential privacy at
● Instantaneous Randomized Response has differential privacy at
= 4 ln(3)
= ln(3)
Differential Privacy of RAPPOR:
Measurable privacy bounds
Each report offers differential privacy with
ε = ln(3)
Attacker’s guess goes from 0.1% → 0.3% in the worst case
Differential privacy even if attacker gets all reports (infinite data!!!)
Also… Base Rate Fallacy prevents attackers from finding needles in
haystacks
Cohorts
Bloom Filter: 2 bits out of 128 — too many false positives
...
user 0xA0FE91B76:
google.com
cohort 2cohort 1 cohort 128
h2
Decoding RAPPOR
From Raw Counts to De-noised Counts
True bit counts, with no noise
De-noised RAPPOR reports
From De-Noised Count to Distribution
True bit counts, with no noise
De-noised RAPPOR reports
google.com:
yahoo.com:
bing.com:
From De-Noised Count to Distribution
Linear Regression:
minX ||B - A X||2
LASSO:
minX (||B - A X||2)2 + λ||X||1
Hybrid:
1. Find support of X via LASSO
2. Solve linear regression to find weights
Deploying RAPPOR
Coverage
Explaining RAPPOR
“Having the cake and eating it too…”
“Seeing the forest without seeing the trees…”
Metaphor for RAPPOR
Microdata: An Individual’s Report
Microdata: An Individual’s Report
Each bit is flipped with
probability
25%
Big Picture Remains!
Google Chrome Privacy White Paper
https://www.google.com/chrome/browser/privacy/whitepaper.html
Phishing and malware protection
Google Chrome includes an optional feature called "Safe Browsing" to help protect you against phishing and malware attacks. This
helps prevent evil-doers from tricking you into sharing personal information with them (“phishing”) or installing malicious software
on your computer (“malware”). The approach used to accomplish this was designed specifically to protect your privacy and is also
used by other popular browsers.
If you'd rather not send any information to Safe Browsing, you can also turn these features off. Please be aware that Chrome will no
longer be able to protect you from websites that try to steal your information or install harmful software if you disable this feature.
We really don't recommend turning it off.
…
If a URL was indeed dangerous, Chrome reports this anonymously to Google to improve Safe Browsing. The data sent is randomized,
constructed in a manner that ensures differential privacy, permitting only monitoring of aggregate statistics that apply to tens of
thousands of users at minimum. The reports are an instance of Randomized Aggregatable Privacy-Preserving Ordinal Responses,
whose full technical details have been published in a technical report and presented at the 2014 ACM Computer and Communications
Security conference. This means that Google cannot infer which website you have visited from this.
Developers’ Uptake
RAPPOR:
Lessons Learned
Growing Pains
● Transitioning from a research prototype to a real product
● Scalability
● Versioning
Communicating Uncertainty
The "Three Commas" Club
Follow-up
- Bassily, Smith, “Local, Private, Efficient Protocols for Succinct
Histograms,” STOC 2015
- Kairouz, Bonawitz, Ramage, “Discrete Distribution Estimation under
Local Privacy”, https://arxiv.org/abs/1602.07387
- Qin et al., “Heavy Hitter Estimation over Set-Valued Data with Local
Differential Privacy”, CCS 2016
Key takeaway points
RAPPOR - locally differentially-private mechanism for reporting of
categorical and string data
● First Internet-scale deployment of differential privacy
● Explainable
● Conservative
● Open-sourced
Apple's On-Device Differential
Privacy
Abhradeep Guha Thakurta, UC Santa Cruz
Apple WWDC, June 2016
References
https://arxiv.org/abs/1709.02753
Phablet
Derp
Photobomb
Woot
Phablet
OMG
Woot
Troll
Prepone
Phablet
awwww
dp
Learning from private data
Learn new (and frequent) words typed
Learning from private data
Learn frequent emojis typed
Apple's On-Device Differential
Privacy: Discovering New Words
Roadmap
1. Private frequency estimation with count-min-sketch
2. Private heavy hitters with puzzle piece algorithm
3. Private heavy hitters with tree histogram protocol
Private Frequency Oracle
Private frequency oracle
Building block for private heavy hitters
𝑑2𝑑1 𝑑 𝑛
All errors within
𝛾 = O( 𝑛 log|𝒮|)
frequency
Words (𝒮)
𝛾
"phablet"
frequency("phablet")
Private frequency oracle:
Design constraints
Computational and communication constraints:
Client side:
Logarithm in size of the domain (|S|) and n
Communication to server:
very few bits
Server-side cost for one query:
size of the domain (|S|) and n
Private frequency oracle:
Design constraints
Computational and communication constraints:
Client side:
size of the domain (|S|) and n
# characters > 3,000
For 8-character words:
size of the domain |S|=3,000^8
number of clients ~ 1B
Efficiently [BS15] ~ n
Our goal ~ O(log |S|)
Private frequency oracle:
Design constraints
Computational and communication constraints:
Client side:
O(log |S|)
Communication to server:
O(1) bits
Server-side cost for one query:
O(log |S|)
Private frequency oracle
A starter solution: Randomized response
𝑑
0 1 0
𝑖
1 0 1
𝑖
Protects ε-differential privacy
(with the right bias)
Randomized response: d′
1 0 0
1 1 0
1 0 1
+ With bias
correction
frequency
All domain elements
Error in each estimate:
Θ( 𝑛 log|𝒮|)
Optimal error under privacy
Private frequency oracle
A starter solution: Randomized response
Private frequency oracle
A starter solution: Randomized response
Computational and communication constraints:
Client side:
O(|S|)
Communication to server:
O(|S|) bits
Server-side cost for one query:
O(1)
1 0 1
𝑖
𝑑
0 01
0 01
0 01
Hash function: ℎ1
Hash function: ℎ2
Hash function: ℎ 𝑘
Number of hash bins: 𝑛
Computation= 𝑂(log|𝒮|)
𝑘 ≈ log|𝒮|
Private frequency oracle
Non-private count-min sketch [CM05]
0 01
0 01
0 01
0 01
1 00
0 11
1
𝑘
1
+
245
127
9123
2132
𝑛
Reducing server computation
Private frequency oracle
Non-private count-min sketch [CM05]
Reducing server computation
1
𝑘
1
Phablet
245
127
9123
2132
𝑛
9146
2212
Frequency estimate:
min (9146, 2212, 2132)
Error in each estimate:
O( 𝑛log|𝒮|)
Server side query cost:
𝑂(log|𝒮|)
𝑘 ≈ log |𝒮|
Private frequency oracle
Non-private count-min sketch [CM05]
"phablet"
Private frequency oracle
Private count-min sketch
𝑑
Making client computation differentially private
0 01
0 01
0 01
1 01
1 00
0 00
𝑘𝜖-diff. private, since 𝑘 pieces of information
Private frequency oracle
Private count-min sketch
𝑑
Theorem: Sampling ensures 𝜖-differential privacy without hurting accuracy,
rather improves it by a factor of 𝑘
0 01 1 00
Private frequency oracle
Private count-min sketch
Reducing client communication
0 01 +1 +1-1
Hadamard transform
Private frequency oracle
Private count-min sketch
Reducing client communication
0 01 +1 +1-1
Hadamard transform
-1 +1
Communication: 𝑂(1) bit
Theorem: Hadamard transform and sampling
do not hurt accuracy
Private frequency oracle
Private count-min sketch
Computational and communication constraints:
Client side:
O(log |S|)
Communication to server:
O(1) bits
Server-side cost for one query:
O(log |S|)
Error in each estimate:
O( 𝑛log|𝒮|)
Roadmap
1. Private frequency estimation with count-min-sketch
2. Private heavy hitters with puzzle piece algorithm
3. Private heavy hitters with tree histogram protocol
Private heavy hitters:
Using the frequency oracle
Private frequency oracle
Private count-min sketch
Domain 𝒮
Too many elements in 𝒮 to search.
Element s in S
Frequency(s)
Find all s in S with
frequency > γ
Roadmap
1. Private frequency estimation with count-min-sketch
2. Private heavy hitters with puzzle piece algorithm
3. Private heavy hitters with tree histogram protocol
Puzzle piece algorithm
(works well in practice, no theoretical guarantees)
[Bassily Nissim Stemmer Thakurta, 2017 and Apple differential privacy team, 2017]
Private heavy hitters
Observation: If a word is frequent, its bigrams are frequent too.
Ph ab le t$ Frequency > 𝛾
Each bi-gram frequency > 𝛾
Private heavy hitters
Natural algorithm: Cartesian product of frequent bi-grams
Sanitized
bi-grams, and the
complete word
ab
ad
ph
ba
ab
ax
le
ab
ab
Position P1 Position P2 Position P3
le
ab
t$
Position P4
Frequent bi-grams
Private heavy hitters
ab
ad
ph
ba
ab
ax
le
ab
ab
Position P1 Position P2 Position P3
le
ab
t$
Position P4
Frequent bi-grams Candidate words
P1 x P2 x P3 x P4
Private frequency oracle
Private count-min sketch
Find frequent
words
Natural algorithm: Cartesian product of frequent bi-grams
Private heavy hitters
Natural algorithm: Cartesian product of frequent bi-grams
Candidate words
P1 x P2 x P3 x P4
Private frequency oracle
Find frequent
words
Combinatorial explosion
In practice, all bi-grams are frequent
Private count-min sketch
Puzzle piece algorithm
Ph ab le t$
≜
h=Hash(Phablet)
Hash: 𝒮 → 1, … , ℓ
Ph ab le t$h h h h
Privatized
bi-grams tagged
with the hash, and
the complete
word
Puzzle piece algorithm: Server side
ab 1
ad 5
Ph 3
ba 4
ab 3
ax 9
le 3
le 7
ab 1
Position P1 Position P2 Position P3
le 1
ab 9
t$ 3
Position P4
Frequent bi-grams tagged with {1, … , ℓ}
Candidate words
P1 x P2 x P3 x P4
Private frequency oracle
Find frequent
words
Combine only matching
bi-grams
Private count-min sketch
Roadmap
1. Private frequency estimation with count-min-sketch
2. Private heavy hitters with puzzle piece algorithm
3. Private heavy hitters with tree histogram protocol
Tree histogram algorithm
(works well in practice + optimal theoretical guarantees)
[Bassily Nissim Stemmer Thakurta, 2017]
Private heavy hitters:
Tree histograms (based on [CM05])
1 0 0
Any string in 𝒮:
log |𝒮| bits
Idea: Construct prefixes of the heavy hitter bit by bit
Private heavy hitters:
Tree histograms
0 1
Private heavy hitters:
Tree histograms
0 1
Level 1: Frequent prefix of length 1
Use private frequency oracle
If a string is a heavy hitter, its prefixes are too.
Private heavy hitters:
Tree histograms
00 01 10 11
Private heavy hitters:
Tree histograms
Level 2: Frequent prefix of length two
Idea: Each level has ≈ 𝑛 heavy hitters
00 01 10 11
Private heavy hitters:
Tree histograms
Computational and communication constraints:
Client side:
O(log |S|)
Communication to server:
O(1) bits
Server-side computation:
O(n log |S|)
Theorem: Finds all heavy hitters with frequency at least
𝑂( 𝑛 log|𝒮|)
Key takeaway points
• Keeping local differential privacy constant:
•One low-noise report is better than many noisy ones
•Weak signal with probability 1 is better than strong signal with small probability
• We can learn the dictionary – at a cost
• Longitudinal privacy remains a challenge
NeurIPS 2017
Microsoft: Discretization of continuous variables
"These guarantees are particularly strong when user’s behavior remains
approximately the same, varies slowly, or varies around a small number of
values over the course of data collection."
Microsoft's deployment
"Our mechanisms have been deployed by
Microsoft across millions of devices ... to protect
users’ privacy while collecting application usage
statistics."
B. Ding, J. Kulkarni, S. Yekhanin, NeurIPS 2017
Microsoft Research Blog, Dec 8, 2017
Privacy in AI @ LinkedIn
• Framework to compute robust, privacy-preserving analytics
• Privacy challenges/design for a large crowdsourced system (LinkedIn Salary)
Analytics & Reporting Products at LinkedIn
Profile View Analytics
137
Content Analytics
Ad Campaign Analytics
All showing
demographics of
members engaging with
the product
• Admit only a small # of predetermined query types
• Querying for the number of member actions, for a specified time period,
together with the top demographic breakdowns
Analytics & Reporting Products at LinkedIn
• Admit only a small # of predetermined query types
• Querying for the number of member actions, for a specified time period,
together with the top demographic breakdowns
E.g., Clicks on a
given adE.g., Title = “Senior
Director”
Analytics & Reporting Products at LinkedIn
Privacy Requirements
• Attacker cannot infer whether a member performed an action
• E.g., click on an article or an ad
• Attacker may use auxiliary knowledge
• E.g., knowledge of attributes associated with the target member (say,
obtained from this member’s LinkedIn profile)
• E.g., knowledge of all other members that performed similar action
Possible Privacy Attacks
141
Targeting:
Senior directors in US, who studied at Cornell
Matches ~16k LinkedIn members
→ over minimum targeting threshold
Demographic breakdown:
Company = X
May match exactly one person
→ can determine whether the person
clicks on the ad or not
Require minimum reporting threshold
Still amenable to attacks
(Refer our ACM CIKM’18 paper for details)
Rounding mechanism
E.g., report incremental of 10
Still amenable to attacks
E.g. using incremental counts over time to
infer individuals’ actions
Need rigorous techniques to preserve member privacy
(not reveal exact aggregate counts)
Key Product Desiderata
• Coverage & Utility
• Data Consistency
• for repeated queries
• over time
• between total and breakdowns
• across entity/action hierarchy
• for top k queries
Problem Statement
Compute robust, reliable analytics in a privacy-
preserving manner, while addressing the product
desiderata such as coverage, utility, and consistency.
Differential Privacy: Random Noise Addition
If ℓ1-sensitivity of f : D → ℝn:
maxD,D′ ||f(D) − f(D′)||1 = s,
then adding Laplacian noise to true output
f(D) + Laplacen(s/ε)
offers ε-differential privacy.
Dwork, McSherry, Nissim, Smith, “Calibrating Noise to Sensitivity in Private Data Analysis”, TCC 2006
PriPeARL: A Framework for Privacy-Preserving Analytics
K. Kenthapadi, T. T. L. Tran, ACM CIKM 2018
145
Pseudo-random noise generation, inspired by differential privacy
● Entity id (e.g., ad
creative/campaign/account)
● Demographic dimension
● Stat type (impressions, clicks)
● Time range
● Fixed secret seed
Uniformly Random
Fraction
● Cryptographic
hash
● Normalize to
(0,1)
Random
Noise
Laplace
Noise
● Fixed ε
True
Count
Noisy
Count
To satisfy consistency
requirements
● Pseudo-random noise → same query has same result over time, avoid
averaging attack.
● For non-canonical queries (e.g., time ranges, aggregate multiple entities)
○ Use the hierarchy and partition into canonical queries
○ Compute noise for each canonical queries and sum up the noisy
counts
System Architecture
Lessons Learned from Deployment (> 1 year)
• Semantic consistency vs. unbiased, unrounded noise
• Suppression of small counts
• Online computation and performance requirements
• Scaling across analytics applications
• Tools for ease of adoption (code/API library, hands-on how-to tutorial) help!
Summary
• Framework to compute robust, privacy-preserving analytics
• Addressing challenges such as preserving member privacy, product coverage, utility,
and data consistency
• Future
• Utility maximization problem given constraints on the ‘privacy loss budget’ per user
• E.g., noise with larger variance to impressions but less noise to clicks (or conversions)
• E.g., more noise to broader time range sub-queries and less noise to granular time range sub-
queries
• Reference: K. Kenthapadi, T. Tran, PriPeARL: A Framework for Privacy-Preserving
Analytics and Reporting at LinkedIn, ACM CIKM 2018.
• https://engineering.linkedin.com/blog/2019/04/privacy-preserving-analytics-and-reporting-at-
linkedin
Acknowledgements
•Team:
• AI/ML: Krishnaram Kenthapadi, Thanh T. L. Tran
• Ad Analytics Product & Engineering: Mark Dietz, Taylor Greason, Ian
Koeppe
• Legal / Security: Sara Harrington, Sharon Lee, Rohit Pitke
•Acknowledgements (in alphabetical order)
• Deepak Agarwal, Igor Perisic, Arun Swami
LinkedIn Salary
Outline
• LinkedIn Salary Overview
• Challenges: Privacy, Modeling
• System Design & Architecture
• Privacy vs. Modeling Tradeoffs
LinkedIn Salary (launched in Nov, 2016)
Salary Collection Flow via Email Targeting
Current Reach (May 2019)
• A few million responses out of several millions of members targeted
• Targeted via emails since early 2016
• Countries: US, CA, UK, DE, IN, …
• Insights available for a large fraction of US monthly active users
Data Privacy Challenges
• Minimize the risk of inferring any one individual’s compensation data
• Protection against data breach
• No single point of failure
Achieved by a combination of
techniques: encryption, access control,
, aggregation,
thresholding
K. Kenthapadi, A. Chudhary, and S.
Ambler, LinkedIn Salary: A System
for Secure Collection and
Presentation of Structured
Compensation Insights to Job
Seekers, IEEE PAC 2017
(arxiv.org/abs/1705.06976)
Modeling Challenges
• Evaluation
• Modeling on de-identified data
• Robustness and stability
• Outlier detection
X. Chen, Y. Liu, L. Zhang, and K.
Kenthapadi, How LinkedIn
Economic Graph Bonds
Information and Product:
Applications in LinkedIn Salary,
KDD 2018
(arxiv.org/abs/1806.09063)
K. Kenthapadi, S. Ambler,
L. Zhang, and D. Agarwal,
Bringing salary transparency to
the world: Computing robust
compensation insights via
LinkedIn Salary, CIKM 2017
(arxiv.org/abs/1703.09845)
Problem Statement
•How do we design LinkedIn Salary system taking into
account the unique privacy and security challenges,
while addressing the product requirements?
Differential Privacy? [Dwork et al, 2006]
• Rich privacy literature (Adam-Worthmann, Samarati-Sweeney, Agrawal-Srikant, …,
Kenthapadi et al, Machanavajjhala et al, Li et al, Dwork et al)
• Limitation of anonymization techniques (as discussed in the first part)
• Worst case sensitivity of quantiles to any one user’s compensation data is
large
•  Large noise to be added, depriving reliability/usefulness
• Need compensation insights on a continual basis
• Theoretical work on applying differential privacy under continual observations
• No practical implementations / applications
• Local differential privacy / Randomized response based approaches (Google’s RAPPOR; Apple’s
iOS differential privacy; Microsoft’s telemetry collection) not applicable
Title Region
$$
User Exp
Designer
SF Bay
Area 100K
User Exp
Designer
SF Bay
Area 115K
... ...
...
Title Region
$$
User Exp
Designer
SF Bay
Area 100K
De-identification Example
Title Region Company Industry Years of
exp
Degree FoS Skills
$$
User Exp
Designer
SF Bay
Area
Google Internet 12 BS Interactive
Media
UX,
Graphics,
...
100K
Title Region Industry
$$
User Exp
Designer
SF Bay
Area
Internet
100K
Title Region Years of
exp $$
User Exp
Designer
SF Bay
Area
10+
100K
Title Region Company Years of
exp $$
User Exp
Designer
SF Bay
Area
Google 10+
100K
#data
points >
threshold?
Yes ⇒ Copy to
Hadoop (HDFS) Note: Original submission stored as encrypted objects.
System
Architecture
Collection
&
Storage
Collection & Storage
• Allow members to submit their compensation info
• Extract member attributes
• E.g., canonical job title, company, region, by invoking LinkedIn standardization services
• Securely store member attributes & compensation data
De-identification
&
Grouping
De-identification & Grouping
• Approach inspired by k-Anonymity [Samarati-Sweeney]
• “Cohort” or “Slice”
• Defined by a combination of attributes
• E.g, “User experience designers in SF Bay Area”
• Contains aggregated compensation entries from corresponding individuals
• No user name, id or any attributes other than those that define the cohort
• A cohort available for offline processing only if it has at least k entries
• Apply LinkedIn standardization software (free-form attribute  canonical version)
before grouping
• Analogous to the generalization step in k-Anonymity
De-identification & Grouping
• Slicing service
• Access member attribute info &
submission identifiers (no
compensation data)
• Generate slices & track #
submissions for each slice
• Preparation service
• Fetch compensation data (using
submission identifiers), associate
with the slice data, copy to HDFS
Insights
&
Modeling
Insights & Modeling
• Salary insight service
• Check whether the member is
eligible
• Give-to-get model
• If yes, show the insights
• Offline workflow
• Consume de-identified HDFS
dataset
• Compute robust compensation
insights
• Outlier detection
• Bayesian smoothing/inference
• Populate the insight key-value
stores
Security
Mechanisms
Security
Mechanisms
• Encryption of
member attributes
& compensation
data using different
sets of keys
• Separation of
processing
• Limiting access to
the keys
Security
Mechanisms
• Key rotation
• No single point of
failure
• Infra security
Preventing Timestamp Join based Attacks
• Inference attack by joining these on timestamp
• De-identified compensation data
• Page view logs (when a member accessed compensation collection web interface)
•  Not desirable to retain the exact timestamp
• Perturb by adding random delay (say, up to 48 hours)
• Modification based on k-Anonymity
• Generalization using a hierarchy of timestamps
• But, need to be incremental
•  Process entries within a cohort in batches of size k
• Generalize to a common timestamp
• Make additional data available only in such incremental batches
Privacy vs Modeling Tradeoffs
• LinkedIn Salary system deployed in production for ~3 years
• Study tradeoffs between privacy guarantees (‘k’) and data available for
computing insights
• Dataset: Compensation submission history from 1.5M LinkedIn members
• Amount of data available vs. minimum threshold, k
• Effect of processing entries in batches of size, k
Amount of
data
available vs.
threshold, k
Percent of
data available
vs. batch size,
k
Median delay
due to
batching vs.
batch size, k
Key takeaway points
• LinkedIn Salary: a new internet application, with
unique privacy/modeling challenges
• Privacy vs. Modeling Tradeoffs
• Potential directions
• Privacy-preserving machine learning models in a practical setting
[e.g., Chaudhuri et al, JMLR 2011; Papernot et al, ICLR 2017]
• Provably private submission of compensation entries?
Beyond Randomized Response
Beyond Randomized Response
• Federated Learning
• DP + Machine Learning
• Encode-Shuffle-Analyze architecture
"Prochlo: Strong Privacy for Analytics in the Crowd"
Bittau et al., SOSP 2017
• Amplification by Shuffling
Federated Learning
"Practical secure aggregation for privacy-preserving machine learning"
Bonawitz, Ivanov, Kreuter, Marcedone, McMahan, Patel, Ramage, Segal,
Seth, ACM CCS 2017
ML and Differential Privacy
"Generalization Implies Privacy" Fallacy
We don’t overfit, therefore
our model cannot possibly
violate privacy.
“Generalization Implies Privacy” Fallacy
Generalization
● average case
● model’s accuracy
Privacy
● worst case
● model’s parameters
“Generalization Implies Privacy” Fallacy
● Examples when it just ain’t so:
○ Person-to-person similarities
○ Support Vector Machines
● Models can be very large
○ Millions of parameters
Somali to English Translation
Somali to English Translation
Somali to English Translation
Somali to English Translation
Somali to English Translation
Maori to English
“Understanding Deep Networks Requires Rethinking
Generalization”, Zhang et al.’17
ML + Differential Privacy
• [DP-SGD] Abadi, Chu, Goodfellow, McMahan, Mironov, Talwar, Zhang,
"Deep Learning with Differential Privacy", ACM CCS 2016
• [PATE] Papernot, Abadi, Erlingson, Goodfellow, Talwar, "Semi-supervised
Knowledge Transfer for Deep Learning from Private Training Data",
ICML 2017
• [PATE] Papernot, Song, Mironov, Raghunathan, Talwar, Erlingson,
"Scalable Private Learning with PATE", ICML 2018
https://github.com/tensorflow/privacy
Statistics + Differential Privacy
Harvard Privacy Tools Project:
Census 2020 and Differential Privacy
Key takeaway points
• Notion of differential privacy is a principled foundation for privacy-
preserving data analyses
• Local differential privacy is a powerful technique appropriate for
Internet-scale telemetry
• Other techniques (thresholding, shuffling) can be combined with
differentially private algorithms or be used in isolation.
References
Differential privacy:
review "A Firm Foundation For Private Data Analysis", C. ACM 2011
by Dwork
book "The Algorithmic Foundations of Differential Privacy"
by Dwork and Roth
References
Google's RAPPOR:
paper "RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response", ACM CCS 2014,
Erlingsson, Pihur, Korolova
blog (https://security.googleblog.com/2014/10/learning-statistics-with-privacy-aided.html)
Apple's implementation:
article "Learning with Privacy at Scale", Apple ML J., Dec 2017
paper "Practical Locally Private Heavy Hitters", NIPS 2017,
by Bassily, Nissim, Stemmer, Thakurta
paper "Privacy Loss in Apple's Implementation of Differential Privacy on MacOS 10.12" by Tang,
Korolova, Bai, Wang, Wang
LinkedIn’s privacy-preserving analytics framework
paper "PriPeARL: A Framework for Privacy-Preserving Analytics and Reporting at LinkedIn", CIKM
2018, Kenthapadi, Tran
blog (https://engineering.linkedin.com/blog/2019/04/privacy-preserving-analytics-and-reporting-
at-linkedin)
LinkedIn Salary:
paper "LinkedIn Salary: A System for Secure Collection and Presentation of Structured
Compensation Insights to Job Seekers", IEEE PAC 2017, Kenthapadi, Chudhary, Ambler
blog (https://engineering.linkedin.com/blog/2017/12/statistical-modeling-for-linkedin-salary)
Fairness Privacy
Transparency Explainability
Related WWW’19 sessions:
1.Tutorial: Designing Equitable Algorithms for the Web
2.Tutorial: Economic Theories of Distributive Justice for Fair Machine Learning
3.Tutorial: Socially Responsible NLP
4.Tutorial: Fairness-aware Machine Learning in Practice
5.Tutorial: Explainable Recommendation and Search
6.Workshop: FATE and Society on the Web
7.Session: Fairness, Credibility, and Search (Wednesday, 10:30 – 12:30)
8.Session: Privacy and Trust (Wednesday, 16:00 – 17:30)
9.Special Track: Designing an Ethical Web (Friday)
Thanks! Questions?
•Tutorial website:
https://sites.google.com/view/www19-privacy-
tutorial
•Feedback most welcome 
• kkenthapadi@linkedin.com, mironov@google.com,
aguhatha@ucsc.edu

More Related Content

What's hot

Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WS...
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WS...Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WS...
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WS...Krishnaram Kenthapadi
 
Machine Learning for Dummies (without mathematics)
Machine Learning for Dummies (without mathematics)Machine Learning for Dummies (without mathematics)
Machine Learning for Dummies (without mathematics)ActiveEon
 
Generative AI Fundamentals - Databricks
Generative AI Fundamentals - DatabricksGenerative AI Fundamentals - Databricks
Generative AI Fundamentals - DatabricksVijayananda Mohire
 
Generative AI - Responsible Path Forward.pdf
Generative AI - Responsible Path Forward.pdfGenerative AI - Responsible Path Forward.pdf
Generative AI - Responsible Path Forward.pdfSaeed Al Dhaheri
 
Privacy preserving machine learning
Privacy preserving machine learningPrivacy preserving machine learning
Privacy preserving machine learningMichał Kuźba
 
Impact of Generative AI in Cybersecurity - How can ISO/IEC 27032 help?
Impact of Generative AI in Cybersecurity - How can ISO/IEC 27032 help?Impact of Generative AI in Cybersecurity - How can ISO/IEC 27032 help?
Impact of Generative AI in Cybersecurity - How can ISO/IEC 27032 help?PECB
 
Generative adversarial networks
Generative adversarial networksGenerative adversarial networks
Generative adversarial networks남주 김
 
Fairness and Privacy in AI/ML Systems
Fairness and Privacy in AI/ML SystemsFairness and Privacy in AI/ML Systems
Fairness and Privacy in AI/ML SystemsKrishnaram Kenthapadi
 
Adversarial Attacks on A.I. Systems — NextCon, Jan 2019
Adversarial Attacks on A.I. Systems — NextCon, Jan 2019Adversarial Attacks on A.I. Systems — NextCon, Jan 2019
Adversarial Attacks on A.I. Systems — NextCon, Jan 2019anant90
 
Interpretable machine learning
Interpretable machine learningInterpretable machine learning
Interpretable machine learningSri Ambati
 
Introduction to machine learning
Introduction to machine learningIntroduction to machine learning
Introduction to machine learningRahul Sahai
 
TEDx Manchester: AI & The Future of Work
TEDx Manchester: AI & The Future of WorkTEDx Manchester: AI & The Future of Work
TEDx Manchester: AI & The Future of WorkVolker Hirsch
 
Introduction to Knowledge Graphs and Semantic AI
Introduction to Knowledge Graphs and Semantic AIIntroduction to Knowledge Graphs and Semantic AI
Introduction to Knowledge Graphs and Semantic AISemantic Web Company
 
Counterfactual Learning for Recommendation
Counterfactual Learning for RecommendationCounterfactual Learning for Recommendation
Counterfactual Learning for RecommendationOlivier Jeunen
 
Generative AI for the rest of us
Generative AI for the rest of usGenerative AI for the rest of us
Generative AI for the rest of usMassimo Ferre'
 

What's hot (20)

Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WS...
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WS...Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WS...
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WS...
 
Machine Learning for Dummies (without mathematics)
Machine Learning for Dummies (without mathematics)Machine Learning for Dummies (without mathematics)
Machine Learning for Dummies (without mathematics)
 
Causal Inference in Marketing
Causal Inference in MarketingCausal Inference in Marketing
Causal Inference in Marketing
 
Generative AI Fundamentals - Databricks
Generative AI Fundamentals - DatabricksGenerative AI Fundamentals - Databricks
Generative AI Fundamentals - Databricks
 
Generative AI - Responsible Path Forward.pdf
Generative AI - Responsible Path Forward.pdfGenerative AI - Responsible Path Forward.pdf
Generative AI - Responsible Path Forward.pdf
 
Privacy preserving machine learning
Privacy preserving machine learningPrivacy preserving machine learning
Privacy preserving machine learning
 
Relationship Between Big Data & AI
Relationship Between Big Data & AIRelationship Between Big Data & AI
Relationship Between Big Data & AI
 
Impact of Generative AI in Cybersecurity - How can ISO/IEC 27032 help?
Impact of Generative AI in Cybersecurity - How can ISO/IEC 27032 help?Impact of Generative AI in Cybersecurity - How can ISO/IEC 27032 help?
Impact of Generative AI in Cybersecurity - How can ISO/IEC 27032 help?
 
Generative adversarial networks
Generative adversarial networksGenerative adversarial networks
Generative adversarial networks
 
Fairness and Privacy in AI/ML Systems
Fairness and Privacy in AI/ML SystemsFairness and Privacy in AI/ML Systems
Fairness and Privacy in AI/ML Systems
 
Adversarial Attacks on A.I. Systems — NextCon, Jan 2019
Adversarial Attacks on A.I. Systems — NextCon, Jan 2019Adversarial Attacks on A.I. Systems — NextCon, Jan 2019
Adversarial Attacks on A.I. Systems — NextCon, Jan 2019
 
Interpretable machine learning
Interpretable machine learningInterpretable machine learning
Interpretable machine learning
 
Introduction to machine learning
Introduction to machine learningIntroduction to machine learning
Introduction to machine learning
 
TEDx Manchester: AI & The Future of Work
TEDx Manchester: AI & The Future of WorkTEDx Manchester: AI & The Future of Work
TEDx Manchester: AI & The Future of Work
 
Introduction to Knowledge Graphs and Semantic AI
Introduction to Knowledge Graphs and Semantic AIIntroduction to Knowledge Graphs and Semantic AI
Introduction to Knowledge Graphs and Semantic AI
 
Counterfactual Learning for Recommendation
Counterfactual Learning for RecommendationCounterfactual Learning for Recommendation
Counterfactual Learning for Recommendation
 
Generative models
Generative modelsGenerative models
Generative models
 
Naive Bayes
Naive Bayes Naive Bayes
Naive Bayes
 
Generative AI for the rest of us
Generative AI for the rest of usGenerative AI for the rest of us
Generative AI for the rest of us
 
Bias in AI
Bias in AIBias in AI
Bias in AI
 

Similar to Privacy-preserving Data Mining in Industry (WWW 2019 Tutorial)

Privacy-preserving Data Mining in Industry (WSDM 2019 Tutorial)
Privacy-preserving Data Mining in Industry (WSDM 2019 Tutorial)Privacy-preserving Data Mining in Industry (WSDM 2019 Tutorial)
Privacy-preserving Data Mining in Industry (WSDM 2019 Tutorial)Krishnaram Kenthapadi
 
Privacy-preserving Data Mining in Industry: Practical Challenges and Lessons ...
Privacy-preserving Data Mining in Industry: Practical Challenges and Lessons ...Privacy-preserving Data Mining in Industry: Practical Challenges and Lessons ...
Privacy-preserving Data Mining in Industry: Practical Challenges and Lessons ...Krishnaram Kenthapadi
 
BROWN BAG TALK WITH MICAH ALTMAN, SOURCES OF BIG DATA FOR SOCIAL SCIENCES
BROWN BAG TALK WITH MICAH ALTMAN, SOURCES OF BIG DATA FOR SOCIAL SCIENCESBROWN BAG TALK WITH MICAH ALTMAN, SOURCES OF BIG DATA FOR SOCIAL SCIENCES
BROWN BAG TALK WITH MICAH ALTMAN, SOURCES OF BIG DATA FOR SOCIAL SCIENCESMicah Altman
 
Fairness, Transparency, and Privacy in AI @ LinkedIn
Fairness, Transparency, and Privacy in AI @ LinkedInFairness, Transparency, and Privacy in AI @ LinkedIn
Fairness, Transparency, and Privacy in AI @ LinkedInKrishnaram Kenthapadi
 
Session 01 designing and scoping a data science project
Session 01 designing and scoping a data science projectSession 01 designing and scoping a data science project
Session 01 designing and scoping a data science projectbodaceacat
 
Session 01 designing and scoping a data science project
Session 01 designing and scoping a data science projectSession 01 designing and scoping a data science project
Session 01 designing and scoping a data science projectSara-Jayne Terp
 
Literature Review: The Role of Signal Processing in Meeting Privacy Challenge...
Literature Review: The Role of Signal Processing in Meeting Privacy Challenge...Literature Review: The Role of Signal Processing in Meeting Privacy Challenge...
Literature Review: The Role of Signal Processing in Meeting Privacy Challenge...Kato Mivule
 
Fairness, Transparency, and Privacy in AI @LinkedIn
Fairness, Transparency, and Privacy in AI @LinkedInFairness, Transparency, and Privacy in AI @LinkedIn
Fairness, Transparency, and Privacy in AI @LinkedInC4Media
 
Ben Shneiderman: Thrill of Discovery
Ben Shneiderman: Thrill of DiscoveryBen Shneiderman: Thrill of Discovery
Ben Shneiderman: Thrill of Discoveryruss9595
 
Contractor-Borner-SNA-SAC
Contractor-Borner-SNA-SACContractor-Borner-SNA-SAC
Contractor-Borner-SNA-SACwebuploader
 
Data Tactics Data Science Brown Bag (April 2014)
Data Tactics Data Science Brown Bag (April 2014)Data Tactics Data Science Brown Bag (April 2014)
Data Tactics Data Science Brown Bag (April 2014)Rich Heimann
 
Data Science and Analytics Brown Bag
Data Science and Analytics Brown BagData Science and Analytics Brown Bag
Data Science and Analytics Brown BagDataTactics
 
Ci2004-10.doc
Ci2004-10.docCi2004-10.doc
Ci2004-10.docbutest
 
Data Science and AI in Biomedicine: The World has Changed
Data Science and AI in Biomedicine: The World has ChangedData Science and AI in Biomedicine: The World has Changed
Data Science and AI in Biomedicine: The World has ChangedPhilip Bourne
 
Characterizing Data and Software for Social Science Research
Characterizing Data and Software for Social Science ResearchCharacterizing Data and Software for Social Science Research
Characterizing Data and Software for Social Science ResearchMicah Altman
 
Immersive Recommendation Workshop, NYC Media Lab'17
Immersive Recommendation Workshop, NYC Media Lab'17Immersive Recommendation Workshop, NYC Media Lab'17
Immersive Recommendation Workshop, NYC Media Lab'17Longqi Yang
 
Don't blindly trust your ML System, it may change your life (Azzurra Ragone, ...
Don't blindly trust your ML System, it may change your life (Azzurra Ragone, ...Don't blindly trust your ML System, it may change your life (Azzurra Ragone, ...
Don't blindly trust your ML System, it may change your life (Azzurra Ragone, ...Data Driven Innovation
 
algorithmic-decisions, fairness, machine learning, provenance, transparency
algorithmic-decisions, fairness, machine learning, provenance, transparencyalgorithmic-decisions, fairness, machine learning, provenance, transparency
algorithmic-decisions, fairness, machine learning, provenance, transparencyPaolo Missier
 
Fairness in Machine Learning
Fairness in Machine LearningFairness in Machine Learning
Fairness in Machine LearningDelip Rao
 

Similar to Privacy-preserving Data Mining in Industry (WWW 2019 Tutorial) (20)

Privacy-preserving Data Mining in Industry (WSDM 2019 Tutorial)
Privacy-preserving Data Mining in Industry (WSDM 2019 Tutorial)Privacy-preserving Data Mining in Industry (WSDM 2019 Tutorial)
Privacy-preserving Data Mining in Industry (WSDM 2019 Tutorial)
 
Privacy-preserving Data Mining in Industry: Practical Challenges and Lessons ...
Privacy-preserving Data Mining in Industry: Practical Challenges and Lessons ...Privacy-preserving Data Mining in Industry: Practical Challenges and Lessons ...
Privacy-preserving Data Mining in Industry: Practical Challenges and Lessons ...
 
BROWN BAG TALK WITH MICAH ALTMAN, SOURCES OF BIG DATA FOR SOCIAL SCIENCES
BROWN BAG TALK WITH MICAH ALTMAN, SOURCES OF BIG DATA FOR SOCIAL SCIENCESBROWN BAG TALK WITH MICAH ALTMAN, SOURCES OF BIG DATA FOR SOCIAL SCIENCES
BROWN BAG TALK WITH MICAH ALTMAN, SOURCES OF BIG DATA FOR SOCIAL SCIENCES
 
Fairness, Transparency, and Privacy in AI @ LinkedIn
Fairness, Transparency, and Privacy in AI @ LinkedInFairness, Transparency, and Privacy in AI @ LinkedIn
Fairness, Transparency, and Privacy in AI @ LinkedIn
 
Session 01 designing and scoping a data science project
Session 01 designing and scoping a data science projectSession 01 designing and scoping a data science project
Session 01 designing and scoping a data science project
 
Session 01 designing and scoping a data science project
Session 01 designing and scoping a data science projectSession 01 designing and scoping a data science project
Session 01 designing and scoping a data science project
 
Why Data Science is a Science
Why Data Science is a ScienceWhy Data Science is a Science
Why Data Science is a Science
 
Literature Review: The Role of Signal Processing in Meeting Privacy Challenge...
Literature Review: The Role of Signal Processing in Meeting Privacy Challenge...Literature Review: The Role of Signal Processing in Meeting Privacy Challenge...
Literature Review: The Role of Signal Processing in Meeting Privacy Challenge...
 
Fairness, Transparency, and Privacy in AI @LinkedIn
Fairness, Transparency, and Privacy in AI @LinkedInFairness, Transparency, and Privacy in AI @LinkedIn
Fairness, Transparency, and Privacy in AI @LinkedIn
 
Ben Shneiderman: Thrill of Discovery
Ben Shneiderman: Thrill of DiscoveryBen Shneiderman: Thrill of Discovery
Ben Shneiderman: Thrill of Discovery
 
Contractor-Borner-SNA-SAC
Contractor-Borner-SNA-SACContractor-Borner-SNA-SAC
Contractor-Borner-SNA-SAC
 
Data Tactics Data Science Brown Bag (April 2014)
Data Tactics Data Science Brown Bag (April 2014)Data Tactics Data Science Brown Bag (April 2014)
Data Tactics Data Science Brown Bag (April 2014)
 
Data Science and Analytics Brown Bag
Data Science and Analytics Brown BagData Science and Analytics Brown Bag
Data Science and Analytics Brown Bag
 
Ci2004-10.doc
Ci2004-10.docCi2004-10.doc
Ci2004-10.doc
 
Data Science and AI in Biomedicine: The World has Changed
Data Science and AI in Biomedicine: The World has ChangedData Science and AI in Biomedicine: The World has Changed
Data Science and AI in Biomedicine: The World has Changed
 
Characterizing Data and Software for Social Science Research
Characterizing Data and Software for Social Science ResearchCharacterizing Data and Software for Social Science Research
Characterizing Data and Software for Social Science Research
 
Immersive Recommendation Workshop, NYC Media Lab'17
Immersive Recommendation Workshop, NYC Media Lab'17Immersive Recommendation Workshop, NYC Media Lab'17
Immersive Recommendation Workshop, NYC Media Lab'17
 
Don't blindly trust your ML System, it may change your life (Azzurra Ragone, ...
Don't blindly trust your ML System, it may change your life (Azzurra Ragone, ...Don't blindly trust your ML System, it may change your life (Azzurra Ragone, ...
Don't blindly trust your ML System, it may change your life (Azzurra Ragone, ...
 
algorithmic-decisions, fairness, machine learning, provenance, transparency
algorithmic-decisions, fairness, machine learning, provenance, transparencyalgorithmic-decisions, fairness, machine learning, provenance, transparency
algorithmic-decisions, fairness, machine learning, provenance, transparency
 
Fairness in Machine Learning
Fairness in Machine LearningFairness in Machine Learning
Fairness in Machine Learning
 

More from Krishnaram Kenthapadi

Responsible AI in Industry: Practical Challenges and Lessons Learned
Responsible AI in Industry: Practical Challenges and Lessons LearnedResponsible AI in Industry: Practical Challenges and Lessons Learned
Responsible AI in Industry: Practical Challenges and Lessons LearnedKrishnaram Kenthapadi
 
Privacy in AI/ML Systems: Practical Challenges and Lessons Learned
Privacy in AI/ML Systems: Practical Challenges and Lessons LearnedPrivacy in AI/ML Systems: Practical Challenges and Lessons Learned
Privacy in AI/ML Systems: Practical Challenges and Lessons LearnedKrishnaram Kenthapadi
 
Explainable AI in Industry (AAAI 2020 Tutorial)
Explainable AI in Industry (AAAI 2020 Tutorial)Explainable AI in Industry (AAAI 2020 Tutorial)
Explainable AI in Industry (AAAI 2020 Tutorial)Krishnaram Kenthapadi
 
Explainable AI in Industry (FAT* 2020 Tutorial)
Explainable AI in Industry (FAT* 2020 Tutorial)Explainable AI in Industry (FAT* 2020 Tutorial)
Explainable AI in Industry (FAT* 2020 Tutorial)Krishnaram Kenthapadi
 
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WW...
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WW...Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WW...
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WW...Krishnaram Kenthapadi
 
Privacy-preserving Analytics and Data Mining at LinkedIn
Privacy-preserving Analytics and Data Mining at LinkedInPrivacy-preserving Analytics and Data Mining at LinkedIn
Privacy-preserving Analytics and Data Mining at LinkedInKrishnaram Kenthapadi
 

More from Krishnaram Kenthapadi (7)

Responsible AI in Industry: Practical Challenges and Lessons Learned
Responsible AI in Industry: Practical Challenges and Lessons LearnedResponsible AI in Industry: Practical Challenges and Lessons Learned
Responsible AI in Industry: Practical Challenges and Lessons Learned
 
Amazon SageMaker Clarify
Amazon SageMaker ClarifyAmazon SageMaker Clarify
Amazon SageMaker Clarify
 
Privacy in AI/ML Systems: Practical Challenges and Lessons Learned
Privacy in AI/ML Systems: Practical Challenges and Lessons LearnedPrivacy in AI/ML Systems: Practical Challenges and Lessons Learned
Privacy in AI/ML Systems: Practical Challenges and Lessons Learned
 
Explainable AI in Industry (AAAI 2020 Tutorial)
Explainable AI in Industry (AAAI 2020 Tutorial)Explainable AI in Industry (AAAI 2020 Tutorial)
Explainable AI in Industry (AAAI 2020 Tutorial)
 
Explainable AI in Industry (FAT* 2020 Tutorial)
Explainable AI in Industry (FAT* 2020 Tutorial)Explainable AI in Industry (FAT* 2020 Tutorial)
Explainable AI in Industry (FAT* 2020 Tutorial)
 
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WW...
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WW...Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WW...
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WW...
 
Privacy-preserving Analytics and Data Mining at LinkedIn
Privacy-preserving Analytics and Data Mining at LinkedInPrivacy-preserving Analytics and Data Mining at LinkedIn
Privacy-preserving Analytics and Data Mining at LinkedIn
 

Privacy-preserving Data Mining in Industry (WWW 2019 Tutorial)

  • 1. Privacy-preserving Data Mining in Industry WWW 2019 Tutorial May 2019 Krishnaram Kenthapadi (AI @ LinkedIn) Ilya Mironov (Google AI) Abhradeep Thakurta (UC Santa Cruz) https://sites.google.com/view/www19-privacy-tutorial
  • 3. Fairness Privacy Transparency Explainability Related WWW’19 sessions: 1.Tutorial: Designing Equitable Algorithms for the Web 2.Tutorial: Economic Theories of Distributive Justice for Fair Machine Learning 3.Tutorial: Socially Responsible NLP 4.Tutorial: Fairness-aware Machine Learning in Practice 5.Tutorial: Explainable Recommendation and Search 6.Workshop: FATE and Society on the Web 7.Session: Fairness, Credibility, and Search (Wednesday, 10:30 – 12:30) 8.Session: Privacy and Trust (Wednesday, 16:00 – 17:30) 9.Special Track: Designing an Ethical Web (Friday)
  • 4. “Privacy by Design” for AI products
  • 5. Outline / Learning Outcomes • Privacy breaches and lessons learned • Evolution of privacy techniques • Differential privacy: definition and techniques • Privacy techniques in practice: Challenges and Lessons Learned • Google’s RAPPOR • Apple’s differential privacy deployment for iOS • Privacy in AI @ LinkedIn (Analytics framework & LinkedIn Salary) • Key Takeaways
  • 6. Privacy: A Historical Perspective Evolution of Privacy Techniques and Privacy Breaches
  • 7. Privacy Breaches and Lessons Learned Attacks on privacy •Governor of Massachusetts •AOL •Netflix •Web browsing data •Facebook •Amazon •Genetic data
  • 8. born July 31, 1945 resident of 02138 Massachusetts Group Insurance Commission (1997): Anonymized medical history of state employees (all hospital visits, diagnosis, prescriptions) Latanya Sweeney (MIT grad student): $20 – Cambridge voter roll William Weld vs Latanya Sweeney
  • 9. 64 %uniquely identifiable with ZIP + birth date + gender (in the US population) Golle, “Revisiting the Uniqueness of Simple Demographics in the US Population”,
  • 11. August 4, 2006: AOL Research publishes anonymized search logs of 650,000 users August 9: New York Times AOL Data Release
  • 12. Attacker's Advantage Auxiliary information Enough to succeed on a small fraction of inputs
  • 14. Oct 2006: Netflix announces Netflix Prize • 10% of their users • average 200 ratings per user Narayanan, Shmatikov (2006): Netflix Prize
  • 15. Deanonymizing Netflix Data Narayanan, Shmatikov, Robust De- anonymization of Large Datasets (How to Break Anonymity of the Netflix Prize Dataset), 2008
  • 16. ● Noam Chomsky in Our Times ● Farenheit 9/11 ● Jesus of Nazareth ● Queer as Folk
  • 17. Key idea: ● Similar intuition as the attack on medical records ● Medical records: Each person can be identified based on a combination of a few attributes ● Web browsing history: Browsing history is unique for each person ● Each person has a distinctive social network  links appearing in one’s feed is unique ● Users likely to visit links in their feed with higher probability than a random user ● “Browsing histories contain tell-tale marks of identity” Su et al, De-anonymizing Web Browsing Data with Social Networks, 2017 De-anonymizing Web Browsing Data with Social Networks
  • 18. Attacker's Advantage Auxiliary information Enough to succeed on a small fraction of inputs High dimensionality
  • 19. Ad targeting: Korolova, “Privacy Violations Using Microtargeted Ads: A Case Study”, PADM Privacy Attacks On Ad Targeting
  • 20. 10 campaigns targeting 1 person (zip code, gender, workplace, alma mater) Korolova, “Privacy Violations Using Microtargeted Ads: A Case Study”, PADM Facebook vs Korolova Age 21 22 23 … 30 Ad Impressions in a week 0 0 8 … 0
  • 21. 10 campaigns targeting 1 person (zip code, gender, workplace, alma mater) Korolova, “Privacy Violations Using Microtargeted Ads: A Case Study”, PADM Facebook vs Korolova Interest A B C … Z Ad Impressions in a week 0 0 8 … 0
  • 22. ● Context: Microtargeted Ads ● Takeaway: Attackers can instrument ad campaigns to identify individual users. ● Two types of attacks: ○ Inference from Impressions ○ Inference from Clicks Facebook vs Korolova: Recap
  • 23. Attacker's Advantage Auxiliary information Enough to succeed on a small fraction of inputs High dimensionality Active
  • 24. Items frequently bought together Bought: A B C D E Z: Customers Who Bought This Item Also Bought Calandrino, Kilzer, Narayanan, Felten, Shmatikov, “You Might Also Like: Privacy Risks of Collaborative Attacking Amazon.com A C D E
  • 25. Attacker's Advantage Auxiliary information Enough to succeed on a small fraction of inputs High dimensionality Active Observant
  • 26. Homer et al., “Resolving individuals contributing trace amounts of DNA to highly complex mixtures using high- density SNP genotyping microarrays”, PLoS Genetics, 2008 Genetic data
  • 28.
  • 29. “In all mixtures, the identification of the presence of a person’s genomic DNA was possible.”
  • 30. Zerhouni, NIH Director: “As a result, the NIH has removed from open-access databases the aggregate results (including P values and genotype counts) for all the GWAS that had been available on NIH sites” … one week later
  • 31. Attacker's Advantage Auxiliary information Enough to succeed on a small fraction of inputs High dimensionality Active Observant Clever
  • 33. Dinur-Nissim 0 1 1 0 1 0 0 0 1 1 0 1Data query: 𝚺 Dinur-Nissim 2003: If error is o(√n), then reconstruction is possible up to n−o(n) ...even if 23.9% of errors are arbitrary [DMT07] ...even with O(n) queries [DY08]
  • 34. Dwork-Naor Tore Dalenius desideratum (aka as “semantic security”): “Access to a statistical database should not enable one to learn anything about an individual that could not be learned without access.” (1977) Dwork-Naor (~2006): If the database teaches us anything, there is always some auxiliary information that breaks Dalenius desideratum.
  • 39. Differential Privacy 39 Databases D and D′ are neighbors if they differ in one person’s data. Differential Privacy: The distribution of the curator’s output M(D) on database D is (nearly) the same as M(D′). CuratorCurator + your data - your data Dwork, McSherry, Nissim, Smith [TCC 2006]
  • 40. ε-Differential Privacy: The distribution of the curator’s output M(D) on database D is (nearly) the same as M(D′). Differential Privacy 40 CuratorCurator Parameter ε quantifies information leakage ∀S: Pr[M(D)∊S] ≤ exp(ε) ∙ Pr[M(D′)∊S]. + your data - your data Dwork, McSherry, Nissim, Smith [TCC 2006]
  • 41. ε-Differential Privacy: The distribution of the curator’s output M(D) on database D is (nearly) the same as M(D′). Differential Privacy 41 CuratorCurator Parameter ε quantifies information leakage ∀S: Pr[M(D)∊S] ≤ exp(ε) ∙ Pr[M(D′)∊S]+𝛿. Parameter 𝛿 gives some slack Dwork, Kenthapadi, McSherry, Mironov, Naor [EUROCRYPT 2006] + your data - your data
  • 42. 42 f(D) f(D′) — bad outcomes — probability with record x — probability without record x “Bad Outcomes” Interpretation
  • 43. ● Prior on databases p ● Observed output O ● Does the database contain record x? 43 Bayesian Interpretation
  • 44. Differential Privacy ● Robustness to auxiliary data ● Post-processing: If M(D) is differentially private, so is f(M(D)). ● Composability: Run two ε-DP mechanisms. Full interaction is 2ε-DP. ● Group privacy: Graceful degradation in the presence of correlated inputs. 44
  • 45. Differential Privacy: Laplace Mechanism Define ℓ1-sensitivity of f: D→ℝn: maxD,D′ ||f(D) − f(D′)||1 < 1, then the Laplace mechanism f(D) + Laplacen(1/ε) offers ε-differential privacy. Dwork, McSherry, Nissim, Smith, “Calibrating Noise to Sensitivity in Private Data Analysis”, TCC 2006
  • 46. Differential Privacy: Gaussian Mechanism If ℓ2-sensitivity of f:D→ℝn: maxD,D′ ||f(D) − f(D′)||2 < 1, then the Gaussian mechanism f(D) + Nn(0, σ2) offers (ε, δ)-differential privacy, where δ = ⅘·exp(-(εσ)2/2). Dwork, Kenthapadi, McSherry, Mironov, Naor, “Our Data, Ourselves”, Eurocrypt 2006
  • 47. What Differential Privacy Isn’t ● Algorithm, architecture, or a rule book ● Secure Computation: what not how ● All-encompassing guarantee: trends may be sensitive too
  • 49. BBC: “Fitness app Strava lights up staff at military bases”
  • 50. Differential Privacy: Takeaway points • Privacy as a notion of stability of randomized algorithms in respect to small perturbations in their input • Worst-case definition • Robust (to auxiliary data, correlated inputs) • Composable • Quantifiable • Concept of a privacy budget • Noise injection
  • 54.
  • 57. Differential Privacy ε-Differential Privacy: The distribution of the output M(D) on database D is (nearly) the same as M(D′) for all adjacent databases D and D′: ∀S: Pr[M(D)∊S] ≤ exp(ε) ∙ Pr[M(D′)∊S].
  • 58. Local Differential Privacy ε-Differential Privacy: The distribution of the output M(D) on database D is (nearly) the same as M(D′) for all adjacent databases D and D′: ∀S: Pr[M(D)∊S] ≤ exp(ε) ∙ Pr[M(D′)∊S].
  • 59. Local-Differentially Private Mechanisms ● Stanley L. Warner, "Randomized response: a survey technique for eliminating evasive answer bias", Journal of American Statistical Association, March 1965. ● Arijit Chaudhuri, Rahul Mukerjee. Randomized Response. Theory and Techniques. 1988.
  • 60. Randomized Response (Warner 1965) Q1: Are you a citizen of the United States? Q2: Are you not a citizen of the United States? 𝜃 - the true fraction of citizens in the sample Answer Q1 Answer Q2 p 1 − p -DP
  • 61. RAPPOR Erlingsson, Pihur, Korolova. "RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response." ACM CCS 2014.
  • 62. RAPPOR: two-level randomized response Can we do repeated surveys of sensitive attributes? — Average of randomized responses will reveal a user’s true answer :-( Solution: Memoize! Re-use the same random answer — Memoization can hurt privacy too! Long, random bit sequence can be a unique tracking ID :-( Solution: Use 2-levels! Randomize the memoized response
  • 63. RAPPOR: two-level randomized response ● Store client value v into bloom filter B using hash functions ● Memoize a Permanent Randomized Response (PRR) B′ ● Report an Instantaneous Randomized Response (IRR) S
  • 64. RAPPOR: two-level randomized response ● Store client value v into bloom filter B using hash functions ● Memoize a Permanent Randomized Response (PRR) B′ ● Report an Instantaneous Randomized Response (IRR) S f = ½ q = ¾ , p = ½
  • 65. RAPPOR: Life of a report Value Bloom Filter PRR IRR “www.google.com”
  • 68. Differential privacy of RAPPOR ● Permanent Randomized Response satisfies differential privacy at ● Instantaneous Randomized Response has differential privacy at = 4 ln(3) = ln(3)
  • 69. Differential Privacy of RAPPOR: Measurable privacy bounds Each report offers differential privacy with ε = ln(3) Attacker’s guess goes from 0.1% → 0.3% in the worst case Differential privacy even if attacker gets all reports (infinite data!!!) Also… Base Rate Fallacy prevents attackers from finding needles in haystacks
  • 70. Cohorts Bloom Filter: 2 bits out of 128 — too many false positives ... user 0xA0FE91B76: google.com cohort 2cohort 1 cohort 128 h2
  • 72. From Raw Counts to De-noised Counts True bit counts, with no noise De-noised RAPPOR reports
  • 73. From De-Noised Count to Distribution True bit counts, with no noise De-noised RAPPOR reports google.com: yahoo.com: bing.com:
  • 74. From De-Noised Count to Distribution Linear Regression: minX ||B - A X||2 LASSO: minX (||B - A X||2)2 + λ||X||1 Hybrid: 1. Find support of X via LASSO 2. Solve linear regression to find weights
  • 77. Explaining RAPPOR “Having the cake and eating it too…” “Seeing the forest without seeing the trees…”
  • 80. Microdata: An Individual’s Report Each bit is flipped with probability 25%
  • 82. Google Chrome Privacy White Paper https://www.google.com/chrome/browser/privacy/whitepaper.html Phishing and malware protection Google Chrome includes an optional feature called "Safe Browsing" to help protect you against phishing and malware attacks. This helps prevent evil-doers from tricking you into sharing personal information with them (“phishing”) or installing malicious software on your computer (“malware”). The approach used to accomplish this was designed specifically to protect your privacy and is also used by other popular browsers. If you'd rather not send any information to Safe Browsing, you can also turn these features off. Please be aware that Chrome will no longer be able to protect you from websites that try to steal your information or install harmful software if you disable this feature. We really don't recommend turning it off. … If a URL was indeed dangerous, Chrome reports this anonymously to Google to improve Safe Browsing. The data sent is randomized, constructed in a manner that ensures differential privacy, permitting only monitoring of aggregate statistics that apply to tens of thousands of users at minimum. The reports are an instance of Randomized Aggregatable Privacy-Preserving Ordinal Responses, whose full technical details have been published in a technical report and presented at the 2014 ACM Computer and Communications Security conference. This means that Google cannot infer which website you have visited from this.
  • 85. Growing Pains ● Transitioning from a research prototype to a real product ● Scalability ● Versioning
  • 88. Follow-up - Bassily, Smith, “Local, Private, Efficient Protocols for Succinct Histograms,” STOC 2015 - Kairouz, Bonawitz, Ramage, “Discrete Distribution Estimation under Local Privacy”, https://arxiv.org/abs/1602.07387 - Qin et al., “Heavy Hitter Estimation over Set-Valued Data with Local Differential Privacy”, CCS 2016
  • 89. Key takeaway points RAPPOR - locally differentially-private mechanism for reporting of categorical and string data ● First Internet-scale deployment of differential privacy ● Explainable ● Conservative ● Open-sourced
  • 90. Apple's On-Device Differential Privacy Abhradeep Guha Thakurta, UC Santa Cruz
  • 94. Learning from private data Learn frequent emojis typed
  • 96. Roadmap 1. Private frequency estimation with count-min-sketch 2. Private heavy hitters with puzzle piece algorithm 3. Private heavy hitters with tree histogram protocol
  • 98. Private frequency oracle Building block for private heavy hitters 𝑑2𝑑1 𝑑 𝑛 All errors within 𝛾 = O( 𝑛 log|𝒮|) frequency Words (𝒮) 𝛾 "phablet" frequency("phablet")
  • 99. Private frequency oracle: Design constraints Computational and communication constraints: Client side: Logarithm in size of the domain (|S|) and n Communication to server: very few bits Server-side cost for one query: size of the domain (|S|) and n
  • 100. Private frequency oracle: Design constraints Computational and communication constraints: Client side: size of the domain (|S|) and n # characters > 3,000 For 8-character words: size of the domain |S|=3,000^8 number of clients ~ 1B Efficiently [BS15] ~ n Our goal ~ O(log |S|)
  • 101. Private frequency oracle: Design constraints Computational and communication constraints: Client side: O(log |S|) Communication to server: O(1) bits Server-side cost for one query: O(log |S|)
  • 102. Private frequency oracle A starter solution: Randomized response 𝑑 0 1 0 𝑖 1 0 1 𝑖 Protects ε-differential privacy (with the right bias) Randomized response: d′
  • 103. 1 0 0 1 1 0 1 0 1 + With bias correction frequency All domain elements Error in each estimate: Θ( 𝑛 log|𝒮|) Optimal error under privacy Private frequency oracle A starter solution: Randomized response
  • 104. Private frequency oracle A starter solution: Randomized response Computational and communication constraints: Client side: O(|S|) Communication to server: O(|S|) bits Server-side cost for one query: O(1) 1 0 1 𝑖
  • 105. 𝑑 0 01 0 01 0 01 Hash function: ℎ1 Hash function: ℎ2 Hash function: ℎ 𝑘 Number of hash bins: 𝑛 Computation= 𝑂(log|𝒮|) 𝑘 ≈ log|𝒮| Private frequency oracle Non-private count-min sketch [CM05]
  • 106. 0 01 0 01 0 01 0 01 1 00 0 11 1 𝑘 1 + 245 127 9123 2132 𝑛 Reducing server computation Private frequency oracle Non-private count-min sketch [CM05]
  • 107. Reducing server computation 1 𝑘 1 Phablet 245 127 9123 2132 𝑛 9146 2212 Frequency estimate: min (9146, 2212, 2132) Error in each estimate: O( 𝑛log|𝒮|) Server side query cost: 𝑂(log|𝒮|) 𝑘 ≈ log |𝒮| Private frequency oracle Non-private count-min sketch [CM05] "phablet"
  • 108. Private frequency oracle Private count-min sketch 𝑑 Making client computation differentially private 0 01 0 01 0 01 1 01 1 00 0 00 𝑘𝜖-diff. private, since 𝑘 pieces of information
  • 109. Private frequency oracle Private count-min sketch 𝑑 Theorem: Sampling ensures 𝜖-differential privacy without hurting accuracy, rather improves it by a factor of 𝑘 0 01 1 00
  • 110. Private frequency oracle Private count-min sketch Reducing client communication 0 01 +1 +1-1 Hadamard transform
  • 111. Private frequency oracle Private count-min sketch Reducing client communication 0 01 +1 +1-1 Hadamard transform -1 +1 Communication: 𝑂(1) bit Theorem: Hadamard transform and sampling do not hurt accuracy
  • 112. Private frequency oracle Private count-min sketch Computational and communication constraints: Client side: O(log |S|) Communication to server: O(1) bits Server-side cost for one query: O(log |S|) Error in each estimate: O( 𝑛log|𝒮|)
  • 113. Roadmap 1. Private frequency estimation with count-min-sketch 2. Private heavy hitters with puzzle piece algorithm 3. Private heavy hitters with tree histogram protocol
  • 114. Private heavy hitters: Using the frequency oracle Private frequency oracle Private count-min sketch Domain 𝒮 Too many elements in 𝒮 to search. Element s in S Frequency(s) Find all s in S with frequency > γ
  • 115. Roadmap 1. Private frequency estimation with count-min-sketch 2. Private heavy hitters with puzzle piece algorithm 3. Private heavy hitters with tree histogram protocol
  • 116. Puzzle piece algorithm (works well in practice, no theoretical guarantees) [Bassily Nissim Stemmer Thakurta, 2017 and Apple differential privacy team, 2017]
  • 117. Private heavy hitters Observation: If a word is frequent, its bigrams are frequent too. Ph ab le t$ Frequency > 𝛾 Each bi-gram frequency > 𝛾
  • 118. Private heavy hitters Natural algorithm: Cartesian product of frequent bi-grams Sanitized bi-grams, and the complete word ab ad ph ba ab ax le ab ab Position P1 Position P2 Position P3 le ab t$ Position P4 Frequent bi-grams
  • 119. Private heavy hitters ab ad ph ba ab ax le ab ab Position P1 Position P2 Position P3 le ab t$ Position P4 Frequent bi-grams Candidate words P1 x P2 x P3 x P4 Private frequency oracle Private count-min sketch Find frequent words Natural algorithm: Cartesian product of frequent bi-grams
  • 120. Private heavy hitters Natural algorithm: Cartesian product of frequent bi-grams Candidate words P1 x P2 x P3 x P4 Private frequency oracle Find frequent words Combinatorial explosion In practice, all bi-grams are frequent Private count-min sketch
  • 121. Puzzle piece algorithm Ph ab le t$ ≜ h=Hash(Phablet) Hash: 𝒮 → 1, … , ℓ Ph ab le t$h h h h Privatized bi-grams tagged with the hash, and the complete word
  • 122. Puzzle piece algorithm: Server side ab 1 ad 5 Ph 3 ba 4 ab 3 ax 9 le 3 le 7 ab 1 Position P1 Position P2 Position P3 le 1 ab 9 t$ 3 Position P4 Frequent bi-grams tagged with {1, … , ℓ} Candidate words P1 x P2 x P3 x P4 Private frequency oracle Find frequent words Combine only matching bi-grams Private count-min sketch
  • 123. Roadmap 1. Private frequency estimation with count-min-sketch 2. Private heavy hitters with puzzle piece algorithm 3. Private heavy hitters with tree histogram protocol
  • 124. Tree histogram algorithm (works well in practice + optimal theoretical guarantees) [Bassily Nissim Stemmer Thakurta, 2017]
  • 125. Private heavy hitters: Tree histograms (based on [CM05]) 1 0 0 Any string in 𝒮: log |𝒮| bits Idea: Construct prefixes of the heavy hitter bit by bit
  • 126. Private heavy hitters: Tree histograms 0 1
  • 127. Private heavy hitters: Tree histograms 0 1 Level 1: Frequent prefix of length 1 Use private frequency oracle If a string is a heavy hitter, its prefixes are too.
  • 128. Private heavy hitters: Tree histograms 00 01 10 11
  • 129. Private heavy hitters: Tree histograms Level 2: Frequent prefix of length two Idea: Each level has ≈ 𝑛 heavy hitters 00 01 10 11
  • 130. Private heavy hitters: Tree histograms Computational and communication constraints: Client side: O(log |S|) Communication to server: O(1) bits Server-side computation: O(n log |S|) Theorem: Finds all heavy hitters with frequency at least 𝑂( 𝑛 log|𝒮|)
  • 131. Key takeaway points • Keeping local differential privacy constant: •One low-noise report is better than many noisy ones •Weak signal with probability 1 is better than strong signal with small probability • We can learn the dictionary – at a cost • Longitudinal privacy remains a challenge
  • 133. Microsoft: Discretization of continuous variables "These guarantees are particularly strong when user’s behavior remains approximately the same, varies slowly, or varies around a small number of values over the course of data collection."
  • 134. Microsoft's deployment "Our mechanisms have been deployed by Microsoft across millions of devices ... to protect users’ privacy while collecting application usage statistics." B. Ding, J. Kulkarni, S. Yekhanin, NeurIPS 2017
  • 135. Microsoft Research Blog, Dec 8, 2017
  • 136. Privacy in AI @ LinkedIn • Framework to compute robust, privacy-preserving analytics • Privacy challenges/design for a large crowdsourced system (LinkedIn Salary)
  • 137. Analytics & Reporting Products at LinkedIn Profile View Analytics 137 Content Analytics Ad Campaign Analytics All showing demographics of members engaging with the product
  • 138. • Admit only a small # of predetermined query types • Querying for the number of member actions, for a specified time period, together with the top demographic breakdowns Analytics & Reporting Products at LinkedIn
  • 139. • Admit only a small # of predetermined query types • Querying for the number of member actions, for a specified time period, together with the top demographic breakdowns E.g., Clicks on a given adE.g., Title = “Senior Director” Analytics & Reporting Products at LinkedIn
  • 140. Privacy Requirements • Attacker cannot infer whether a member performed an action • E.g., click on an article or an ad • Attacker may use auxiliary knowledge • E.g., knowledge of attributes associated with the target member (say, obtained from this member’s LinkedIn profile) • E.g., knowledge of all other members that performed similar action
  • 141. Possible Privacy Attacks 141 Targeting: Senior directors in US, who studied at Cornell Matches ~16k LinkedIn members → over minimum targeting threshold Demographic breakdown: Company = X May match exactly one person → can determine whether the person clicks on the ad or not Require minimum reporting threshold Still amenable to attacks (Refer our ACM CIKM’18 paper for details) Rounding mechanism E.g., report incremental of 10 Still amenable to attacks E.g. using incremental counts over time to infer individuals’ actions Need rigorous techniques to preserve member privacy (not reveal exact aggregate counts)
  • 142. Key Product Desiderata • Coverage & Utility • Data Consistency • for repeated queries • over time • between total and breakdowns • across entity/action hierarchy • for top k queries
  • 143. Problem Statement Compute robust, reliable analytics in a privacy- preserving manner, while addressing the product desiderata such as coverage, utility, and consistency.
  • 144. Differential Privacy: Random Noise Addition If ℓ1-sensitivity of f : D → ℝn: maxD,D′ ||f(D) − f(D′)||1 = s, then adding Laplacian noise to true output f(D) + Laplacen(s/ε) offers ε-differential privacy. Dwork, McSherry, Nissim, Smith, “Calibrating Noise to Sensitivity in Private Data Analysis”, TCC 2006
  • 145. PriPeARL: A Framework for Privacy-Preserving Analytics K. Kenthapadi, T. T. L. Tran, ACM CIKM 2018 145 Pseudo-random noise generation, inspired by differential privacy ● Entity id (e.g., ad creative/campaign/account) ● Demographic dimension ● Stat type (impressions, clicks) ● Time range ● Fixed secret seed Uniformly Random Fraction ● Cryptographic hash ● Normalize to (0,1) Random Noise Laplace Noise ● Fixed ε True Count Noisy Count To satisfy consistency requirements ● Pseudo-random noise → same query has same result over time, avoid averaging attack. ● For non-canonical queries (e.g., time ranges, aggregate multiple entities) ○ Use the hierarchy and partition into canonical queries ○ Compute noise for each canonical queries and sum up the noisy counts
  • 147. Lessons Learned from Deployment (> 1 year) • Semantic consistency vs. unbiased, unrounded noise • Suppression of small counts • Online computation and performance requirements • Scaling across analytics applications • Tools for ease of adoption (code/API library, hands-on how-to tutorial) help!
  • 148. Summary • Framework to compute robust, privacy-preserving analytics • Addressing challenges such as preserving member privacy, product coverage, utility, and data consistency • Future • Utility maximization problem given constraints on the ‘privacy loss budget’ per user • E.g., noise with larger variance to impressions but less noise to clicks (or conversions) • E.g., more noise to broader time range sub-queries and less noise to granular time range sub- queries • Reference: K. Kenthapadi, T. Tran, PriPeARL: A Framework for Privacy-Preserving Analytics and Reporting at LinkedIn, ACM CIKM 2018. • https://engineering.linkedin.com/blog/2019/04/privacy-preserving-analytics-and-reporting-at- linkedin
  • 149. Acknowledgements •Team: • AI/ML: Krishnaram Kenthapadi, Thanh T. L. Tran • Ad Analytics Product & Engineering: Mark Dietz, Taylor Greason, Ian Koeppe • Legal / Security: Sara Harrington, Sharon Lee, Rohit Pitke •Acknowledgements (in alphabetical order) • Deepak Agarwal, Igor Perisic, Arun Swami
  • 151. Outline • LinkedIn Salary Overview • Challenges: Privacy, Modeling • System Design & Architecture • Privacy vs. Modeling Tradeoffs
  • 152. LinkedIn Salary (launched in Nov, 2016)
  • 153. Salary Collection Flow via Email Targeting
  • 154. Current Reach (May 2019) • A few million responses out of several millions of members targeted • Targeted via emails since early 2016 • Countries: US, CA, UK, DE, IN, … • Insights available for a large fraction of US monthly active users
  • 155. Data Privacy Challenges • Minimize the risk of inferring any one individual’s compensation data • Protection against data breach • No single point of failure Achieved by a combination of techniques: encryption, access control, , aggregation, thresholding K. Kenthapadi, A. Chudhary, and S. Ambler, LinkedIn Salary: A System for Secure Collection and Presentation of Structured Compensation Insights to Job Seekers, IEEE PAC 2017 (arxiv.org/abs/1705.06976)
  • 156. Modeling Challenges • Evaluation • Modeling on de-identified data • Robustness and stability • Outlier detection X. Chen, Y. Liu, L. Zhang, and K. Kenthapadi, How LinkedIn Economic Graph Bonds Information and Product: Applications in LinkedIn Salary, KDD 2018 (arxiv.org/abs/1806.09063) K. Kenthapadi, S. Ambler, L. Zhang, and D. Agarwal, Bringing salary transparency to the world: Computing robust compensation insights via LinkedIn Salary, CIKM 2017 (arxiv.org/abs/1703.09845)
  • 157. Problem Statement •How do we design LinkedIn Salary system taking into account the unique privacy and security challenges, while addressing the product requirements?
  • 158. Differential Privacy? [Dwork et al, 2006] • Rich privacy literature (Adam-Worthmann, Samarati-Sweeney, Agrawal-Srikant, …, Kenthapadi et al, Machanavajjhala et al, Li et al, Dwork et al) • Limitation of anonymization techniques (as discussed in the first part) • Worst case sensitivity of quantiles to any one user’s compensation data is large •  Large noise to be added, depriving reliability/usefulness • Need compensation insights on a continual basis • Theoretical work on applying differential privacy under continual observations • No practical implementations / applications • Local differential privacy / Randomized response based approaches (Google’s RAPPOR; Apple’s iOS differential privacy; Microsoft’s telemetry collection) not applicable
  • 159. Title Region $$ User Exp Designer SF Bay Area 100K User Exp Designer SF Bay Area 115K ... ... ... Title Region $$ User Exp Designer SF Bay Area 100K De-identification Example Title Region Company Industry Years of exp Degree FoS Skills $$ User Exp Designer SF Bay Area Google Internet 12 BS Interactive Media UX, Graphics, ... 100K Title Region Industry $$ User Exp Designer SF Bay Area Internet 100K Title Region Years of exp $$ User Exp Designer SF Bay Area 10+ 100K Title Region Company Years of exp $$ User Exp Designer SF Bay Area Google 10+ 100K #data points > threshold? Yes ⇒ Copy to Hadoop (HDFS) Note: Original submission stored as encrypted objects.
  • 162. Collection & Storage • Allow members to submit their compensation info • Extract member attributes • E.g., canonical job title, company, region, by invoking LinkedIn standardization services • Securely store member attributes & compensation data
  • 164. De-identification & Grouping • Approach inspired by k-Anonymity [Samarati-Sweeney] • “Cohort” or “Slice” • Defined by a combination of attributes • E.g, “User experience designers in SF Bay Area” • Contains aggregated compensation entries from corresponding individuals • No user name, id or any attributes other than those that define the cohort • A cohort available for offline processing only if it has at least k entries • Apply LinkedIn standardization software (free-form attribute  canonical version) before grouping • Analogous to the generalization step in k-Anonymity
  • 165. De-identification & Grouping • Slicing service • Access member attribute info & submission identifiers (no compensation data) • Generate slices & track # submissions for each slice • Preparation service • Fetch compensation data (using submission identifiers), associate with the slice data, copy to HDFS
  • 167. Insights & Modeling • Salary insight service • Check whether the member is eligible • Give-to-get model • If yes, show the insights • Offline workflow • Consume de-identified HDFS dataset • Compute robust compensation insights • Outlier detection • Bayesian smoothing/inference • Populate the insight key-value stores
  • 169. Security Mechanisms • Encryption of member attributes & compensation data using different sets of keys • Separation of processing • Limiting access to the keys
  • 170. Security Mechanisms • Key rotation • No single point of failure • Infra security
  • 171. Preventing Timestamp Join based Attacks • Inference attack by joining these on timestamp • De-identified compensation data • Page view logs (when a member accessed compensation collection web interface) •  Not desirable to retain the exact timestamp • Perturb by adding random delay (say, up to 48 hours) • Modification based on k-Anonymity • Generalization using a hierarchy of timestamps • But, need to be incremental •  Process entries within a cohort in batches of size k • Generalize to a common timestamp • Make additional data available only in such incremental batches
  • 172. Privacy vs Modeling Tradeoffs • LinkedIn Salary system deployed in production for ~3 years • Study tradeoffs between privacy guarantees (‘k’) and data available for computing insights • Dataset: Compensation submission history from 1.5M LinkedIn members • Amount of data available vs. minimum threshold, k • Effect of processing entries in batches of size, k
  • 175. Median delay due to batching vs. batch size, k
  • 176. Key takeaway points • LinkedIn Salary: a new internet application, with unique privacy/modeling challenges • Privacy vs. Modeling Tradeoffs • Potential directions • Privacy-preserving machine learning models in a practical setting [e.g., Chaudhuri et al, JMLR 2011; Papernot et al, ICLR 2017] • Provably private submission of compensation entries?
  • 178. Beyond Randomized Response • Federated Learning • DP + Machine Learning • Encode-Shuffle-Analyze architecture "Prochlo: Strong Privacy for Analytics in the Crowd" Bittau et al., SOSP 2017 • Amplification by Shuffling
  • 179. Federated Learning "Practical secure aggregation for privacy-preserving machine learning" Bonawitz, Ivanov, Kreuter, Marcedone, McMahan, Patel, Ramage, Segal, Seth, ACM CCS 2017
  • 180.
  • 182. "Generalization Implies Privacy" Fallacy We don’t overfit, therefore our model cannot possibly violate privacy.
  • 183. “Generalization Implies Privacy” Fallacy Generalization ● average case ● model’s accuracy Privacy ● worst case ● model’s parameters
  • 184. “Generalization Implies Privacy” Fallacy ● Examples when it just ain’t so: ○ Person-to-person similarities ○ Support Vector Machines ● Models can be very large ○ Millions of parameters
  • 185. Somali to English Translation
  • 186. Somali to English Translation
  • 187. Somali to English Translation
  • 188. Somali to English Translation
  • 189. Somali to English Translation
  • 191. “Understanding Deep Networks Requires Rethinking Generalization”, Zhang et al.’17
  • 192. ML + Differential Privacy • [DP-SGD] Abadi, Chu, Goodfellow, McMahan, Mironov, Talwar, Zhang, "Deep Learning with Differential Privacy", ACM CCS 2016 • [PATE] Papernot, Abadi, Erlingson, Goodfellow, Talwar, "Semi-supervised Knowledge Transfer for Deep Learning from Private Training Data", ICML 2017 • [PATE] Papernot, Song, Mironov, Raghunathan, Talwar, Erlingson, "Scalable Private Learning with PATE", ICML 2018 https://github.com/tensorflow/privacy
  • 193. Statistics + Differential Privacy Harvard Privacy Tools Project:
  • 194. Census 2020 and Differential Privacy
  • 195. Key takeaway points • Notion of differential privacy is a principled foundation for privacy- preserving data analyses • Local differential privacy is a powerful technique appropriate for Internet-scale telemetry • Other techniques (thresholding, shuffling) can be combined with differentially private algorithms or be used in isolation.
  • 196. References Differential privacy: review "A Firm Foundation For Private Data Analysis", C. ACM 2011 by Dwork book "The Algorithmic Foundations of Differential Privacy" by Dwork and Roth
  • 197. References Google's RAPPOR: paper "RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response", ACM CCS 2014, Erlingsson, Pihur, Korolova blog (https://security.googleblog.com/2014/10/learning-statistics-with-privacy-aided.html) Apple's implementation: article "Learning with Privacy at Scale", Apple ML J., Dec 2017 paper "Practical Locally Private Heavy Hitters", NIPS 2017, by Bassily, Nissim, Stemmer, Thakurta paper "Privacy Loss in Apple's Implementation of Differential Privacy on MacOS 10.12" by Tang, Korolova, Bai, Wang, Wang LinkedIn’s privacy-preserving analytics framework paper "PriPeARL: A Framework for Privacy-Preserving Analytics and Reporting at LinkedIn", CIKM 2018, Kenthapadi, Tran blog (https://engineering.linkedin.com/blog/2019/04/privacy-preserving-analytics-and-reporting- at-linkedin) LinkedIn Salary: paper "LinkedIn Salary: A System for Secure Collection and Presentation of Structured Compensation Insights to Job Seekers", IEEE PAC 2017, Kenthapadi, Chudhary, Ambler blog (https://engineering.linkedin.com/blog/2017/12/statistical-modeling-for-linkedin-salary)
  • 198. Fairness Privacy Transparency Explainability Related WWW’19 sessions: 1.Tutorial: Designing Equitable Algorithms for the Web 2.Tutorial: Economic Theories of Distributive Justice for Fair Machine Learning 3.Tutorial: Socially Responsible NLP 4.Tutorial: Fairness-aware Machine Learning in Practice 5.Tutorial: Explainable Recommendation and Search 6.Workshop: FATE and Society on the Web 7.Session: Fairness, Credibility, and Search (Wednesday, 10:30 – 12:30) 8.Session: Privacy and Trust (Wednesday, 16:00 – 17:30) 9.Special Track: Designing an Ethical Web (Friday)
  • 199. Thanks! Questions? •Tutorial website: https://sites.google.com/view/www19-privacy- tutorial •Feedback most welcome  • kkenthapadi@linkedin.com, mironov@google.com, aguhatha@ucsc.edu