SlideShare a Scribd company logo

Fairness and Privacy in AI/ML Systems

Download as PPTX, PDF
K
Krishnaram Kenthapadi

The document discusses fairness and privacy challenges in AI/ML systems, highlighting the consequences of algorithmic biases and the importance of integrating fairness and privacy by design. It illustrates initiatives taken by LinkedIn and AWS to promote diversity in talent search, along with the development of privacy-preserving analytics tools like Pripearl and secure salary data systems. The document emphasizes the need for collaboration among stakeholders to ensure ethical AI applications while addressing both business metrics and societal impacts.

Internet
1 of 67
Downloaded 61 times
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
Fairness and Privacy in AI/ML Systems Krishnaram Kenthapadi Amazon AWS AI LinkedIn-MSR-IISc workshop on Fairness & Ethics in ML January 2020 https://www.csa.iisc.ac.in/fate.htm
Massachusetts Group Insurance Commission (1997): Anonymized medical history of state employees William Weld vs Latanya Sweeney Latanya Sweeney (MIT grad student): $20 – Cambridge voter roll born July 31, 1945 resident of 02138
Uniquely identifiable with ZIP + birth date + gender (in the US population) Golle, β€œRevisiting the Uniqueness of Simple Demographics in the US Population”, WPES 2006
The Coded Gaze [Joy Buolamwini 2016] Face detection software: Fails for some darker faces https://www.youtube.com/watch?v=KB9sI9rY3cA
β€’ Facial analysis software: Higher accuracy for light skinned men β€’ Error rates for dark skinned women: 20% - 34% Gender Shades [Joy Buolamwini & Timnit Gebru, 2018]
Fairness and Privacy in AI/ML Systems
β€’ Ethical challenges posed by AI systems β€’ Inherent biases present in society β€’ Reflected in training data β€’ AI/ML models prone to amplifying such biases Algorithmic Bias
Laws against Discrimination Immigration Reform and Control Act Citizenship Rehabilitation Act of 1973; Americans with Disabilities Act of 1990 Disability status Civil Rights Act of 1964 Race Age Discrimination in Employment Act of 1967 Age Equal Pay Act of 1963; Civil Rights Act of 1964 Sex And more...
Fairness Privacy Transparency Explainability
Fairness and Privacy by Design” for AI products β€œ
AI @ Scale Case Studies @ LinkedIn* Fairness Privacy Reflections *Work done while at LinkedIn
12Β© 2019 Amazon Web Services, Inc. or its affiliates. All rights reserved | Our mission at AWS Put machine learning in the hands of every developer
13Β© 2019 Amazon Web Services, Inc. or its affiliates. All rights reserved | The AWS ML Stack Broadest and most complete set of Machine Learning capabilities VISION SPEECH TEXT SEARCH NEW CHATBOTS PERSONALIZATION FORECASTING FRAUD NEW DEVELOPMENT NEW CONTACT CENTERS NEW Amazon SageMaker Ground Truth Augmented AI SageMaker Neo Built-in algorithms SageMaker Notebooks NEW SageMaker Experiments NEW Model tuning SageMaker Debugger NEW SageMaker Autopilot NEW Model hosting SageMaker Model Monitor NEW Deep Learning AMIs & Containers GPUs & CPUs Elastic Inference Inferentia FPGA Amazon Rekognition Amazon Polly Amazon Transcribe +Medical Amazon Comprehend +Medical Amazon Translate Amazon Lex Amazon Personalize Amazon Forecast Amazon Fraud Detector Amazon CodeGuru AI SERVICES ML SERVICES ML FRAMEWORKS & INFRASTRUCTURE Amazon Textract Amazon Kendra Contact Lens For Amazon Connect SageMaker Studio IDE NEW NEW NEW NEW NEW
LinkedIn operates the largest professional network on the Internet Tell your story 645M+ members 30M+ companies are represented on LinkedIn 90K+ schools listed (high school & college) 35K+ skills listed 20M+ open jobs on LinkedIn Jobs 280B Feed updates
How AI is transforming LinkedIn’s ecosystem 2 PB+ Data processed nearline and offline per day 25 B+ Parameters in Machine Learning models 200+ Machine Learning A/B experiments per week Contributors Advertising Revenue Confirmed Hires
Fairness in AI @ LinkedIn Fairness-aware Talent Search Ranking
Guiding Principle: β€œDiversity by Design”
Insights to Identify Diverse Talent Pools Representative Talent Search Results Diversity Learning Curriculum β€œDiversity by Design” in LinkedIn’s Talent Solutions
Plan for Diversity
Plan for Diversity
Identify Diverse Talent Pools
Inclusive Job Descriptions / Recruiter Outreach
Representative Ranking for Talent Search S. C. Geyik, S. Ambler, K. Kenthapadi, Fairness- Aware Ranking in Search & Recommendation Systems with Application to LinkedIn Talent Search, KDD’19. [Microsoft’s AI/ML conference (MLADS’18). Distinguished Contribution Award] Building Representative Talent Search at LinkedIn (LinkedIn engineering blog)
Intuition for Measuring and Achieving Representativeness Ideal: Top ranked results should follow a desired distribution on gender/age/… E.g., same distribution as the underlying talent pool Inspired by β€œEqual Opportunity” definition [Hardt et al, NIPS’16] Defined measures (skew, divergence) based on this intuition
Desired Proportions within the Attribute of Interest Compute the proportions of the values of the attribute (e.g., gender, gender-age combination) amongst the set of qualified candidates β€œQualified candidates” = Set of candidates that match the search query criteria Retrieved by LinkedIn’s Galene search engine Desired proportions could also be obtained based on legal mandate / voluntary commitment
Measuring (Lack of) Representativeness Skew@k (Logarithmic) ratio of the proportion of candidates having a given attribute value among the top k ranked results to the corresponding desired proportion Variants: MinSkew: Minimum over all attribute values MaxSkew: Maximum over all attribute values Normalized Discounted Cumulative Skew Normalized Discounted Cumulative KL-divergence
Fairness-aware Reranking Algorithm (Simplified) Partition the set of potential candidates into different buckets for each attribute value Rank the candidates in each bucket according to the scores assigned by the machine-learned model Merge the ranked lists, balancing the representation requirements and the selection of highest scored candidates Algorithmic variants based on how we choose the next attribute
Architecture
Validating Our Approach Gender Representativeness Over 95% of all searches are representative compared to the qualified population of the search Business Metrics A/B test over LinkedIn Recruiter users for two weeks No significant change in business metrics (e.g., # InMails sent or accepted) Ramped to 100% of LinkedIn Recruiter users worldwide
Lessons learned β€’ Post-processing approach desirable β€’ Model agnostic β€’ Scalable across different model choices for our application β€’ Acts as a β€œfail-safe” β€’ Robust to application-specific business logic β€’ Easier to incorporate as part of existing systems β€’ Build a stand-alone service or component for post-processing β€’ No significant modifications to the existing components β€’ Complementary to efforts to reduce bias from training data & during model training
Engineering for Fairness in AI Lifecycle Problem Formation Dataset Construction Algorithm Selection Training Process Testing Process Deployment Feedback Is an algorithm an ethical solution to our problem? Does our data include enough minority samples? Are there missing/biased features? Do we need to apply debiasing algorithms to preprocess our data? Do we need to include fairness constraints in the function? Have we evaluated the model using relevant fairness metrics? Are we deploying our model on a population that we did not train/test on? Does the model encourage feedback loops that can produce increasingly unfair outcomes? Credit: K. Browne & J. Draper
Engineering for Fairness in AI Lifecycle S.Vasudevan, K. Kenthapadi, FairScale: A Scalable Framework for Measuring Fairness in AI Applications, 2019
FairScale System Architecture [Vasudevan & Kenthapadi, 2019] β€’ Flexibility of Use (Platform agnostic) β€’ Ad-hoc exploratory analyses β€’ Deployment in offline workflows β€’ Integration with ML Frameworks β€’ Scalability β€’ Diverse fairness metrics β€’ Conventional fairness metrics β€’ Benefit metrics β€’ Statistical tests
Fairness-aware Experimentation [Saint-Jacques & Sepehri, KDD’19 Social Impact Workshop] Imagine LinkedIn has 10 members. Each of them has 1 session a day. A new product increases sessions by +1 session per member on average. Both of these are +1 session / member on average! One is much more unequal than the other. We want to catch that.
Acknowledgements LinkedIn Talent Solutions Diversity team, Hire & Careers AI team, Anti-abuse AI team, Data Science Applied Research team Special thanks to Deepak Agarwal, Parvez Ahammad, Stuart Ambler, Kinjal Basu, Jenelle Bray, Erik Buchanan, Bee-Chung Chen, Fei Chen, Patrick Cheung, Gil Cottle, Cyrus DiCiccio, Patrick Driscoll, Carlos Faham, Nadia Fawaz, Priyanka Gariba, Meg Garlinghouse, Sahin Cem Geyik, Gurwinder Gulati, Rob Hallman, Sara Harrington, Joshua Hartman, Daniel Hewlett, Nicolas Kim, Rachel Kumar, Monica Lewis, Nicole Li, Heloise Logan, Stephen Lynch, Divyakumar Menghani, Varun Mithal, Arashpreet Singh Mor, Tanvi Motwani, Preetam Nandy, Lei Ni, Nitin Panjwani, Igor Perisic, Hema Raghavan, Romer Rosales, Guillaume Saint-Jacques, Badrul Sarwar, Amir Sepehri, Arun Swami, Ram Swaminathan, Grace Tang, Ketan Thakkar, Sriram Vasudevan, Janardhanan Vembunarayanan, James Verbus, Xin Wang, Hinkmond Wong, Ya Xu, Lin Yang, Yang Yang, Chenhui Zhai, Liang Zhang, Yani Zhang
Privacy in AI @ LinkedIn PriPeARL: Framework to compute robust, privacy-preserving analytics
Analytics & Reporting Products at LinkedIn Profile View Analytics 37 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 Analytics & Reporting Products at LinkedIn E.g., Title = β€œSenior Director” E.g., Clicks on a given ad
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 (say, by creating fake accounts)
Possible Privacy Attacks 41 Targeting: Senior directors in US, who studied at Cornell Matches ~16k LinkedIn members β†’ over minimum targeting threshold Demographic breakdown: Company = X May match exactly one person β†’ can determine whether the person clicks on the ad or not Require minimum reporting threshold Attacker could create fake profiles! E.g. if threshold is 10, create 9 fake profiles that all click. 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)
Problem Statement Compute robust, reliable analytics in a privacy- preserving manner, while addressing the product needs.
Differential Privacy
Curator Defining Privacy
Defining Privacy 45 CuratorCurator + your data - your data
Differential Privacy 46 Databases D and Dβ€² 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(Dβ€²). Curator + your data - your data Dwork, McSherry, Nissim, Smith [TCC 2006] Curator
(Ξ΅, 𝛿)-Differential Privacy: The distribution of the curator’s output M(D) on database D is (nearly) the same as M(Dβ€²). Differential Privacy 47 Curator Parameter Ξ΅ quantifies information leakage βˆ€S: Pr[M(D)∊S] ≀ exp(Ξ΅) βˆ™ Pr[M(Dβ€²)∊S]＋𝛿.Curator Parameter 𝛿 gives some slack Dwork, Kenthapadi, McSherry, Mironov, Naor [EUROCRYPT 2006] + your data - your data Dwork, McSherry, Nissim, Smith [TCC 2006]
Differential Privacy: Random Noise Addition If β„“1-sensitivity of f : D β†’ ℝn: maxD,Dβ€² ||f(D) βˆ’ f(Dβ€²)||1 = s, then adding Laplacian noise to true output f(D) + Laplacen(s/Ξ΅) offers (Ξ΅,0)-differential privacy. Dwork, McSherry, Nissim, Smith [TCC 2006]
PriPeARL: A Framework for Privacy-Preserving Analytics K. Kenthapadi, T. T. L. Tran, ACM CIKM 2018 49 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 Ξ΅ True Count Noisy Count To satisfy consistency requirements ● Pseudo-random noise β†’ same query has same result over time, avoid averaging attack. ● For non-canonical queries (e.g., time ranges, aggregate multiple entities) β—‹ Use the hierarchy and partition into canonical queries β—‹ Compute noise for each canonical queries and sum up the noisy counts
PriPeARL 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! Having a few entry points (all analytics apps built over Pinot)  wider adoption
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 β€˜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.
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 Deepak Agarwal, Igor Perisic, Arun Swami
LinkedIn Salary
LinkedIn Salary (launched in Nov, 2016)
Data Privacy Challenges Minimize the risk of inferring any one individual’s compensation data Protection against data breach No single point of failure
Problem Statement How do we design LinkedIn Salary system taking into account the unique privacy and security challenges, while addressing the product requirements? 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)
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 β‡’ Copy to Hadoop (HDFS) Note: Original submission stored as encrypted objects.
System Architecture
Acknowledgements Team: AI/ML: Krishnaram Kenthapadi, Stuart Ambler, Xi Chen, Yiqun Liu, Parul Jain, Liang Zhang, Ganesh Venkataraman, Tim Converse, Deepak Agarwal Application Engineering: Ahsan Chudhary, Alan Yang, Alex Navasardyan, Brandyn Bennett, Hrishikesh S, Jim Tao, Juan Pablo Lomeli Diaz, Patrick Schutz, Ricky Yan, Lu Zheng, Stephanie Chou, Joseph Florencio, Santosh Kumar Kancha, Anthony Duerr Product: Ryan Sandler, Keren Baruch Other teams (UED, Marketing, BizOps, Analytics, Testing, Voice of Members, Security, …): Julie Kuang, Phil Bunge, Prateek Janardhan, Fiona Li, Bharath Shetty, Sunil Mahadeshwar, Cory Scott, Tushar Dalvi, and team Acknowledgements David Freeman, Ashish Gupta, David Hardtke, Rong Rong, Ram
Beyond Accuracy Performance and Cost Fairness and Bias Transparency and Explainability Privacy Security Safety Robustness
Fairness, Explainability & Privacy: Opportunities
Fairness in ML Application specific challenges Conversational AI systems: Unique bias/fairness/ethics considerations E.g., Hate speech, Complex failure modes Beyond protected categories, e.g., accent, dialect Entire ecosystem (e.g., including apps such as Alexa skills) Two-sided markets: e.g., fairness to buyers and to sellers, or to content consumers and producers Fairness in advertising (externalities) Tools for ensuring fairness (measuring & mitigating bias) in AI lifecycle Pre-processing (representative datasets; modifying features/labels) ML model training with fairness constraints Post-processing Experimentation & Post-deployment
Explainability in ML Actionable explanations Balance between explanations & model secrecy Robustness of explanations to failure modes (Interaction between ML components) Application-specific challenges Conversational AI systems: contextual explanations Gradation of explanations Tools for explanations across AI lifecycle Pre & post-deployment for ML models Model developer vs. End user focused
Privacy in ML Privacy-preserving model training, robust against adversarial membership inference attacks Privacy for highly sensitive data: model training & analytics using secure enclaves, homomorphic encryption, federated learning / on- device learning, or a hybrid Privacy-preserving transfer learning (broadly, privacy-preserving mechanisms for data marketplaces)
Reflections β€œFairness and Privacy by Design” when building AI products Collaboration/consensus across key stakeholders NYT / WSJ / ProPublica / ToI / The Hindu test :)
Thanks! Questions? S. C. Geyik, S. Ambler, K. Kenthapadi, Fairness-Aware Ranking in Search & Recommendation Systems with Application to LinkedIn Talent Search, KDD’19 [Microsoft’s AI/ML conference (MLADS’18). Distinguished Contribution Award] K. Kenthapadi, T. T. L. Tran, PriPeARL: A Framework for Privacy-Preserving Analytics and Reporting at LinkedIn, CIKM’18 K. Kenthapadi, A. Chudhary, S. Ambler, LinkedIn Salary, IEEE Symposium on Privacy-Aware Computing (PAC), 2017 [Related: our KDD’18 & CIKM’17 (Best Case Studies Paper Award) papers] Our tutorials on privacy, on fairness, and on explainability in industry at KDD/WSDM/WWW/FAccT/AAAI (combining experiences at Apple, Facebook, Google, LinkedIn, Microsoft)

More Related Content

PPTX
Responsible AI in Industry: Practical Challenges and Lessons Learned
Krishnaram Kenthapadi
Β 
PPTX
Explainable AI in Industry (KDD 2019 Tutorial)
Krishnaram Kenthapadi
Β 
PDF
Loan approval prediction based on machine learning approach
Eslam Nader
Β 
PDF
An introduction to Recommender Systems
David Zibriczky
Β 
PDF
Fairness and Bias in Machine Learning
Surya Dutta
Β 
PPTX
Loan prediction
Akshay Jadhav
Β 
PPTX
Explainable AI in Industry (WWW 2020 Tutorial)
Krishnaram Kenthapadi
Β 
PDF
Responsible AI
Data Con LA
Β 
Responsible AI in Industry: Practical Challenges and Lessons Learned
Krishnaram Kenthapadi
Β 
Explainable AI in Industry (KDD 2019 Tutorial)
Krishnaram Kenthapadi
Β 
Loan approval prediction based on machine learning approach
Eslam Nader
Β 
An introduction to Recommender Systems
David Zibriczky
Β 
Fairness and Bias in Machine Learning
Surya Dutta
Β 
Loan prediction
Akshay Jadhav
Β 
Explainable AI in Industry (WWW 2020 Tutorial)
Krishnaram Kenthapadi
Β 
Responsible AI
Data Con LA
Β 

What's hot (20)

PDF
Explainable AI (XAI)
Manojkumar Parmar
Β 
PDF
Big Data [sorry] & Data Science: What Does a Data Scientist Do?
Data Science London
Β 
PPTX
Case Study: Loan default prediction
ALTEN Calsoft Labs
Β 
PDF
Ibm machine learning for z os
Cuneyt Goksu
Β 
PPTX
Machine Learning Tutorial | Machine Learning Basics | Machine Learning Algori...
Simplilearn
Β 
PPTX
Responsible AI in Industry (ICML 2021 Tutorial)
Krishnaram Kenthapadi
Β 
PPTX
Feature Selection in Machine Learning
Upekha Vandebona
Β 
PPTX
Anomaly Detection Technique
Chakrit Phain
Β 
PPTX
Ethical Issues in Machine Learning Algorithms. (Part 1)
Vladimir Kanchev
Β 
PDF
Synthetic data generation
Sandeep Joshi
Β 
PPTX
Object detection with deep learning
Sushant Shrivastava
Β 
PPTX
Responsible AI
Anand Rao
Β 
PDF
Synthetic data generation for machine learning
QuantUniversity
Β 
PDF
Explainable AI (XAI) - A Perspective
Saurabh Kaushik
Β 
PDF
Data Catalog as the Platform for Data Intelligence
Alation
Β 
PDF
Data platform architecture
Sudheer Kondla
Β 
PDF
Machine Learning - Object Detection and Classification
Vikas Jain
Β 
PPTX
Data science & data scientist
VijayMohan Vasu
Β 
PPTX
Supervised and Unsupervised Learning In Machine Learning | Machine Learning T...
Simplilearn
Β 
PDF
AIF360 - Trusted and Fair AI
Animesh Singh
Β 
Explainable AI (XAI)
Manojkumar Parmar
Β 
Big Data [sorry] & Data Science: What Does a Data Scientist Do?
Data Science London
Β 
Case Study: Loan default prediction
ALTEN Calsoft Labs
Β 
Ibm machine learning for z os
Cuneyt Goksu
Β 
Machine Learning Tutorial | Machine Learning Basics | Machine Learning Algori...
Simplilearn
Β 
Responsible AI in Industry (ICML 2021 Tutorial)
Krishnaram Kenthapadi
Β 
Feature Selection in Machine Learning
Upekha Vandebona
Β 
Anomaly Detection Technique
Chakrit Phain
Β 
Ethical Issues in Machine Learning Algorithms. (Part 1)
Vladimir Kanchev
Β 
Synthetic data generation
Sandeep Joshi
Β 
Object detection with deep learning
Sushant Shrivastava
Β 
Responsible AI
Anand Rao
Β 
Synthetic data generation for machine learning
QuantUniversity
Β 
Explainable AI (XAI) - A Perspective
Saurabh Kaushik
Β 
Data Catalog as the Platform for Data Intelligence
Alation
Β 
Data platform architecture
Sudheer Kondla
Β 
Machine Learning - Object Detection and Classification
Vikas Jain
Β 
Data science & data scientist
VijayMohan Vasu
Β 
Supervised and Unsupervised Learning In Machine Learning | Machine Learning T...
Simplilearn
Β 
AIF360 - Trusted and Fair AI
Animesh Singh
Β 
Ad

Similar to Fairness and Privacy in AI/ML Systems (20)

PPTX
Fairness and Privacy in AI/ML Systems
Krishnaram Kenthapadi
Β 
PDF
Intro to machine learning for web folks @ BlendWebMix
Louis Dorard
Β 
PDF
Responsible Data Use in AI - core tech pillars
Sofus MacskΓ‘ssy
Β 
PPTX
Machine Learning: Addressing the Disillusionment to Bring Actual Business Ben...
Jon Mead
Β 
PPTX
Data Intelligence: How the Amalgamation of Data, Science, and Technology is C...
Caserta
Β 
PPTX
Responsible AI in Industry: Practical Challenges and Lessons Learned
Krishnaram Kenthapadi
Β 
PDF
Trusted, Transparent and Fair AI using Open Source
Animesh Singh
Β 
PDF
Projects
Learnbay Datascience
Β 
PPTX
[DSC Adria 23] Muthu Ramachandran AI Ethics Framework for Generative AI such ...
DataScienceConferenc1
Β 
PPTX
Explainability for Natural Language Processing
Yunyao Li
Β 
PPTX
Microsoft for Startups program, designed to help new ventures succeed in comp...
NoorUlHaq47
Β 
PDF
State of the Market - Data Quality in 2023
RTTS
Β 
PDF
Explainability for Natural Language Processing
Yunyao Li
Β 
PPTX
Automation, Analytics, and Artificial Intelligence - Panel
AnandSRao1962
Β 
PPTX
Explainability for Natural Language Processing
Yunyao Li
Β 
PDF
CRM Options for Enterprise Nonprofits - Blackbaud CRM Solutions
Heller Consulting
Β 
PDF
Level Up Your Skills to Lead IT Projects Successfully
Kaali Dass PMP, PhD.
Β 
PPTX
Responsible AI in Industry (Tutorials at AAAI 2021, FAccT 2021, and WWW 2021)
Krishnaram Kenthapadi
Β 
PPTX
[DSC Europe 22] AI Ethics and AI Quality By Design - Muthu Ramachandran
DataScienceConferenc1
Β 
PDF
Loras College 2016 Business Analytics Symposium Keynote
Rich Clayton
Β 
Fairness and Privacy in AI/ML Systems
Krishnaram Kenthapadi
Β 
Intro to machine learning for web folks @ BlendWebMix
Louis Dorard
Β 
Responsible Data Use in AI - core tech pillars
Sofus MacskΓ‘ssy
Β 
Machine Learning: Addressing the Disillusionment to Bring Actual Business Ben...
Jon Mead
Β 
Data Intelligence: How the Amalgamation of Data, Science, and Technology is C...
Caserta
Β 
Responsible AI in Industry: Practical Challenges and Lessons Learned
Krishnaram Kenthapadi
Β 
Trusted, Transparent and Fair AI using Open Source
Animesh Singh
Β 
Projects
Learnbay Datascience
Β 
[DSC Adria 23] Muthu Ramachandran AI Ethics Framework for Generative AI such ...
DataScienceConferenc1
Β 
Explainability for Natural Language Processing
Yunyao Li
Β 
Microsoft for Startups program, designed to help new ventures succeed in comp...
NoorUlHaq47
Β 
State of the Market - Data Quality in 2023
RTTS
Β 
Explainability for Natural Language Processing
Yunyao Li
Β 
Automation, Analytics, and Artificial Intelligence - Panel
AnandSRao1962
Β 
Explainability for Natural Language Processing
Yunyao Li
Β 
CRM Options for Enterprise Nonprofits - Blackbaud CRM Solutions
Heller Consulting
Β 
Level Up Your Skills to Lead IT Projects Successfully
Kaali Dass PMP, PhD.
Β 
Responsible AI in Industry (Tutorials at AAAI 2021, FAccT 2021, and WWW 2021)
Krishnaram Kenthapadi
Β 
[DSC Europe 22] AI Ethics and AI Quality By Design - Muthu Ramachandran
DataScienceConferenc1
Β 
Loras College 2016 Business Analytics Symposium Keynote
Rich Clayton
Β 
Ad

More from Krishnaram Kenthapadi (12)

PPTX
Amazon SageMaker Clarify
Krishnaram Kenthapadi
Β 
PPTX
Privacy in AI/ML Systems: Practical Challenges and Lessons Learned
Krishnaram Kenthapadi
Β 
PPTX
Explainable AI in Industry (AAAI 2020 Tutorial)
Krishnaram Kenthapadi
Β 
PPTX
Explainable AI in Industry (FAT* 2020 Tutorial)
Krishnaram Kenthapadi
Β 
PDF
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (KD...
Krishnaram Kenthapadi
Β 
PPTX
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WW...
Krishnaram Kenthapadi
Β 
PPTX
Privacy-preserving Data Mining in Industry (WWW 2019 Tutorial)
Krishnaram Kenthapadi
Β 
PPTX
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WS...
Krishnaram Kenthapadi
Β 
PPTX
Privacy-preserving Data Mining in Industry (WSDM 2019 Tutorial)
Krishnaram Kenthapadi
Β 
PPTX
Fairness, Transparency, and Privacy in AI @ LinkedIn
Krishnaram Kenthapadi
Β 
PPTX
Privacy-preserving Analytics and Data Mining at LinkedIn
Krishnaram Kenthapadi
Β 
PPTX
Privacy-preserving Data Mining in Industry: Practical Challenges and Lessons ...
Krishnaram Kenthapadi
Β 
Amazon SageMaker Clarify
Krishnaram Kenthapadi
Β 
Privacy in AI/ML Systems: Practical Challenges and Lessons Learned
Krishnaram Kenthapadi
Β 
Explainable AI in Industry (AAAI 2020 Tutorial)
Krishnaram Kenthapadi
Β 
Explainable AI in Industry (FAT* 2020 Tutorial)
Krishnaram Kenthapadi
Β 
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (KD...
Krishnaram Kenthapadi
Β 
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WW...
Krishnaram Kenthapadi
Β 
Privacy-preserving Data Mining in Industry (WWW 2019 Tutorial)
Krishnaram Kenthapadi
Β 
Fairness-aware Machine Learning: Practical Challenges and Lessons Learned (WS...
Krishnaram Kenthapadi
Β 
Privacy-preserving Data Mining in Industry (WSDM 2019 Tutorial)
Krishnaram Kenthapadi
Β 
Fairness, Transparency, and Privacy in AI @ LinkedIn
Krishnaram Kenthapadi
Β 
Privacy-preserving Analytics and Data Mining at LinkedIn
Krishnaram Kenthapadi
Β 
Privacy-preserving Data Mining in Industry: Practical Challenges and Lessons ...
Krishnaram Kenthapadi
Β 

Recently uploaded (20)

PPTX
Internet___Basics___Styled_ presentation
poonam62133
Β 
PPTX
innovation process that make everything different.pptx
SoumyadipPaul18
Β 
PPTX
CHE NAA, , b,mn,mblblblbljb jb jlb ,j , ,C PPT.pptx
sumodmjohn3
Β 
PPTX
SAP Ariba Sourcing PPT for learning material
NKKumar16
Β 
PDF
APNIC Update, presented at PHNOG 2025 by Shane Hermoso
APNIC
Β 
PPTX
June-4-Sermon-Powerpoint.pptx USE THIS FOR YOUR MOTIVATION
KuyaRogie
Β 
PDF
πŸ’° π”πŠπ“πˆ πŠπ„πŒπ„ππ€ππ†π€π πŠπˆππ„π‘πŸ’πƒ π‡π€π‘πˆ 𝐈𝐍𝐈 πŸπŸŽπŸπŸ“ πŸ’°
GRAB
Β 
PPT
Design_with_Watersergyerge45hrbgre4top (1).ppt
DiogoRibeiro730590
Β 
PDF
Decoding a Decade: 10 Years of Applied CTI Discipline
Andreas Sfakianakis
Β 
PPTX
introduction about ICD -10 & ICD-11 ppt.pptx
naveenithkrishnan
Β 
PPTX
Introduction about ICD -10 and ICD11 on 5.8.25.pptx
naveenithkrishnan
Β 
PDF
RPKI Status Update, presented by Makito Lay at IDNOG 10
APNIC
Β 
PDF
Unit-1 introduction to cyber security discuss about how to secure a system
shivangisharma636228
Β 
PPTX
Job_Card_System_Styled_lorem_ipsum_.pptx
Jeffkigen
Β 
PPTX
INTERNET------BASICS-------UPDATED PPT PRESENTATION
poonam62133
Β 
PPTX
Module 1 - Cyber Law and Ethics 101.pptx
anantyakhrisna1
Β 
PDF
Testing WebRTC applications at scale.pdf
Giacomo Vacca
Β 
PPT
tcp ip networks nd ip layering assotred slides
pakadamfdp
Β 
PDF
The New Creative Director: How AI Tools for Social Media Content Creation Are...
SageTitans
Β 
PDF
WebRTC in SignalWire - troubleshooting media negotiation
Giacomo Vacca
Β 
Internet___Basics___Styled_ presentation
poonam62133
Β 
innovation process that make everything different.pptx
SoumyadipPaul18
Β 
CHE NAA, , b,mn,mblblblbljb jb jlb ,j , ,C PPT.pptx
sumodmjohn3
Β 
SAP Ariba Sourcing PPT for learning material
NKKumar16
Β 
APNIC Update, presented at PHNOG 2025 by Shane Hermoso
APNIC
Β 
June-4-Sermon-Powerpoint.pptx USE THIS FOR YOUR MOTIVATION
KuyaRogie
Β 
πŸ’° π”πŠπ“πˆ πŠπ„πŒπ„ππ€ππ†π€π πŠπˆππ„π‘πŸ’πƒ π‡π€π‘πˆ 𝐈𝐍𝐈 πŸπŸŽπŸπŸ“ πŸ’°
GRAB
Β 
Design_with_Watersergyerge45hrbgre4top (1).ppt
DiogoRibeiro730590
Β 
Decoding a Decade: 10 Years of Applied CTI Discipline
Andreas Sfakianakis
Β 
introduction about ICD -10 & ICD-11 ppt.pptx
naveenithkrishnan
Β 
Introduction about ICD -10 and ICD11 on 5.8.25.pptx
naveenithkrishnan
Β 
RPKI Status Update, presented by Makito Lay at IDNOG 10
APNIC
Β 
Unit-1 introduction to cyber security discuss about how to secure a system
shivangisharma636228
Β 
Job_Card_System_Styled_lorem_ipsum_.pptx
Jeffkigen
Β 
INTERNET------BASICS-------UPDATED PPT PRESENTATION
poonam62133
Β 
Module 1 - Cyber Law and Ethics 101.pptx
anantyakhrisna1
Β 
Testing WebRTC applications at scale.pdf
Giacomo Vacca
Β 
tcp ip networks nd ip layering assotred slides
pakadamfdp
Β 
The New Creative Director: How AI Tools for Social Media Content Creation Are...
SageTitans
Β 
WebRTC in SignalWire - troubleshooting media negotiation
Giacomo Vacca
Β 

Fairness and Privacy in AI/ML Systems

  • 1. Fairness and Privacy in AI/ML Systems Krishnaram Kenthapadi Amazon AWS AI LinkedIn-MSR-IISc workshop on Fairness & Ethics in ML January 2020 https://www.csa.iisc.ac.in/fate.htm
  • 2. Massachusetts Group Insurance Commission (1997): Anonymized medical history of state employees William Weld vs Latanya Sweeney Latanya Sweeney (MIT grad student): $20 – Cambridge voter roll born July 31, 1945 resident of 02138
  • 3. Uniquely identifiable with ZIP + birth date + gender (in the US population) Golle, β€œRevisiting the Uniqueness of Simple Demographics in the US Population”, WPES 2006
  • 4. The Coded Gaze [Joy Buolamwini 2016] Face detection software: Fails for some darker faces https://www.youtube.com/watch?v=KB9sI9rY3cA
  • 5. β€’ Facial analysis software: Higher accuracy for light skinned men β€’ Error rates for dark skinned women: 20% - 34% Gender Shades [Joy Buolamwini & Timnit Gebru, 2018]
  • 7. β€’ Ethical challenges posed by AI systems β€’ Inherent biases present in society β€’ Reflected in training data β€’ AI/ML models prone to amplifying such biases Algorithmic Bias
  • 8. Laws against Discrimination Immigration Reform and Control Act Citizenship Rehabilitation Act of 1973; Americans with Disabilities Act of 1990 Disability status Civil Rights Act of 1964 Race Age Discrimination in Employment Act of 1967 Age Equal Pay Act of 1963; Civil Rights Act of 1964 Sex And more...
  • 11. AI @ Scale Case Studies @ LinkedIn* Fairness Privacy Reflections *Work done while at LinkedIn
  • 12. 12Β© 2019 Amazon Web Services, Inc. or its affiliates. All rights reserved | Our mission at AWS Put machine learning in the hands of every developer
  • 13. 13Β© 2019 Amazon Web Services, Inc. or its affiliates. All rights reserved | The AWS ML Stack Broadest and most complete set of Machine Learning capabilities VISION SPEECH TEXT SEARCH NEW CHATBOTS PERSONALIZATION FORECASTING FRAUD NEW DEVELOPMENT NEW CONTACT CENTERS NEW Amazon SageMaker Ground Truth Augmented AI SageMaker Neo Built-in algorithms SageMaker Notebooks NEW SageMaker Experiments NEW Model tuning SageMaker Debugger NEW SageMaker Autopilot NEW Model hosting SageMaker Model Monitor NEW Deep Learning AMIs & Containers GPUs & CPUs Elastic Inference Inferentia FPGA Amazon Rekognition Amazon Polly Amazon Transcribe +Medical Amazon Comprehend +Medical Amazon Translate Amazon Lex Amazon Personalize Amazon Forecast Amazon Fraud Detector Amazon CodeGuru AI SERVICES ML SERVICES ML FRAMEWORKS & INFRASTRUCTURE Amazon Textract Amazon Kendra Contact Lens For Amazon Connect SageMaker Studio IDE NEW NEW NEW NEW NEW
  • 14. LinkedIn operates the largest professional network on the Internet Tell your story 645M+ members 30M+ companies are represented on LinkedIn 90K+ schools listed (high school & college) 35K+ skills listed 20M+ open jobs on LinkedIn Jobs 280B Feed updates
  • 15. How AI is transforming LinkedIn’s ecosystem 2 PB+ Data processed nearline and offline per day 25 B+ Parameters in Machine Learning models 200+ Machine Learning A/B experiments per week Contributors Advertising Revenue Confirmed Hires
  • 18. Insights to Identify Diverse Talent Pools Representative Talent Search Results Diversity Learning Curriculum β€œDiversity by Design” in LinkedIn’s Talent Solutions
  • 22. Inclusive Job Descriptions / Recruiter Outreach
  • 23. Representative Ranking for Talent Search S. C. Geyik, S. Ambler, K. Kenthapadi, Fairness- Aware Ranking in Search & Recommendation Systems with Application to LinkedIn Talent Search, KDD’19. [Microsoft’s AI/ML conference (MLADS’18). Distinguished Contribution Award] Building Representative Talent Search at LinkedIn (LinkedIn engineering blog)
  • 24. Intuition for Measuring and Achieving Representativeness Ideal: Top ranked results should follow a desired distribution on gender/age/… E.g., same distribution as the underlying talent pool Inspired by β€œEqual Opportunity” definition [Hardt et al, NIPS’16] Defined measures (skew, divergence) based on this intuition
  • 25. Desired Proportions within the Attribute of Interest Compute the proportions of the values of the attribute (e.g., gender, gender-age combination) amongst the set of qualified candidates β€œQualified candidates” = Set of candidates that match the search query criteria Retrieved by LinkedIn’s Galene search engine Desired proportions could also be obtained based on legal mandate / voluntary commitment
  • 26. Measuring (Lack of) Representativeness Skew@k (Logarithmic) ratio of the proportion of candidates having a given attribute value among the top k ranked results to the corresponding desired proportion Variants: MinSkew: Minimum over all attribute values MaxSkew: Maximum over all attribute values Normalized Discounted Cumulative Skew Normalized Discounted Cumulative KL-divergence
  • 27. Fairness-aware Reranking Algorithm (Simplified) Partition the set of potential candidates into different buckets for each attribute value Rank the candidates in each bucket according to the scores assigned by the machine-learned model Merge the ranked lists, balancing the representation requirements and the selection of highest scored candidates Algorithmic variants based on how we choose the next attribute
  • 29. Validating Our Approach Gender Representativeness Over 95% of all searches are representative compared to the qualified population of the search Business Metrics A/B test over LinkedIn Recruiter users for two weeks No significant change in business metrics (e.g., # InMails sent or accepted) Ramped to 100% of LinkedIn Recruiter users worldwide
  • 30. Lessons learned β€’ Post-processing approach desirable β€’ Model agnostic β€’ Scalable across different model choices for our application β€’ Acts as a β€œfail-safe” β€’ Robust to application-specific business logic β€’ Easier to incorporate as part of existing systems β€’ Build a stand-alone service or component for post-processing β€’ No significant modifications to the existing components β€’ Complementary to efforts to reduce bias from training data & during model training
  • 31. Engineering for Fairness in AI Lifecycle Problem Formation Dataset Construction Algorithm Selection Training Process Testing Process Deployment Feedback Is an algorithm an ethical solution to our problem? Does our data include enough minority samples? Are there missing/biased features? Do we need to apply debiasing algorithms to preprocess our data? Do we need to include fairness constraints in the function? Have we evaluated the model using relevant fairness metrics? Are we deploying our model on a population that we did not train/test on? Does the model encourage feedback loops that can produce increasingly unfair outcomes? Credit: K. Browne & J. Draper
  • 32. Engineering for Fairness in AI Lifecycle S.Vasudevan, K. Kenthapadi, FairScale: A Scalable Framework for Measuring Fairness in AI Applications, 2019
  • 33. FairScale System Architecture [Vasudevan & Kenthapadi, 2019] β€’ Flexibility of Use (Platform agnostic) β€’ Ad-hoc exploratory analyses β€’ Deployment in offline workflows β€’ Integration with ML Frameworks β€’ Scalability β€’ Diverse fairness metrics β€’ Conventional fairness metrics β€’ Benefit metrics β€’ Statistical tests
  • 34. Fairness-aware Experimentation [Saint-Jacques & Sepehri, KDD’19 Social Impact Workshop] Imagine LinkedIn has 10 members. Each of them has 1 session a day. A new product increases sessions by +1 session per member on average. Both of these are +1 session / member on average! One is much more unequal than the other. We want to catch that.
  • 35. Acknowledgements LinkedIn Talent Solutions Diversity team, Hire & Careers AI team, Anti-abuse AI team, Data Science Applied Research team Special thanks to Deepak Agarwal, Parvez Ahammad, Stuart Ambler, Kinjal Basu, Jenelle Bray, Erik Buchanan, Bee-Chung Chen, Fei Chen, Patrick Cheung, Gil Cottle, Cyrus DiCiccio, Patrick Driscoll, Carlos Faham, Nadia Fawaz, Priyanka Gariba, Meg Garlinghouse, Sahin Cem Geyik, Gurwinder Gulati, Rob Hallman, Sara Harrington, Joshua Hartman, Daniel Hewlett, Nicolas Kim, Rachel Kumar, Monica Lewis, Nicole Li, Heloise Logan, Stephen Lynch, Divyakumar Menghani, Varun Mithal, Arashpreet Singh Mor, Tanvi Motwani, Preetam Nandy, Lei Ni, Nitin Panjwani, Igor Perisic, Hema Raghavan, Romer Rosales, Guillaume Saint-Jacques, Badrul Sarwar, Amir Sepehri, Arun Swami, Ram Swaminathan, Grace Tang, Ketan Thakkar, Sriram Vasudevan, Janardhanan Vembunarayanan, James Verbus, Xin Wang, Hinkmond Wong, Ya Xu, Lin Yang, Yang Yang, Chenhui Zhai, Liang Zhang, Yani Zhang
  • 36. Privacy in AI @ LinkedIn PriPeARL: Framework to compute robust, privacy-preserving analytics
  • 37. Analytics & Reporting Products at LinkedIn Profile View Analytics 37 Content Analytics Ad Campaign Analytics All showing demographics of members engaging with the product
  • 38. 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
  • 39. 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 E.g., Title = β€œSenior Director” E.g., Clicks on a given ad
  • 40. 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 (say, by creating fake accounts)
  • 41. Possible Privacy Attacks 41 Targeting: Senior directors in US, who studied at Cornell Matches ~16k LinkedIn members β†’ over minimum targeting threshold Demographic breakdown: Company = X May match exactly one person β†’ can determine whether the person clicks on the ad or not Require minimum reporting threshold Attacker could create fake profiles! E.g. if threshold is 10, create 9 fake profiles that all click. 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)
  • 42. Problem Statement Compute robust, reliable analytics in a privacy- preserving manner, while addressing the product needs.
  • 46. Differential Privacy 46 Databases D and Dβ€² 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(Dβ€²). Curator + your data - your data Dwork, McSherry, Nissim, Smith [TCC 2006] Curator
  • 47. (Ξ΅, 𝛿)-Differential Privacy: The distribution of the curator’s output M(D) on database D is (nearly) the same as M(Dβ€²). Differential Privacy 47 Curator Parameter Ξ΅ quantifies information leakage βˆ€S: Pr[M(D)∊S] ≀ exp(Ξ΅) βˆ™ Pr[M(Dβ€²)∊S]＋𝛿.Curator Parameter 𝛿 gives some slack Dwork, Kenthapadi, McSherry, Mironov, Naor [EUROCRYPT 2006] + your data - your data Dwork, McSherry, Nissim, Smith [TCC 2006]
  • 48. Differential Privacy: Random Noise Addition If β„“1-sensitivity of f : D β†’ ℝn: maxD,Dβ€² ||f(D) βˆ’ f(Dβ€²)||1 = s, then adding Laplacian noise to true output f(D) + Laplacen(s/Ξ΅) offers (Ξ΅,0)-differential privacy. Dwork, McSherry, Nissim, Smith [TCC 2006]
  • 49. PriPeARL: A Framework for Privacy-Preserving Analytics K. Kenthapadi, T. T. L. Tran, ACM CIKM 2018 49 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 Ξ΅ True Count Noisy Count To satisfy consistency requirements ● Pseudo-random noise β†’ same query has same result over time, avoid averaging attack. ● For non-canonical queries (e.g., time ranges, aggregate multiple entities) β—‹ Use the hierarchy and partition into canonical queries β—‹ Compute noise for each canonical queries and sum up the noisy counts
  • 51. 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! Having a few entry points (all analytics apps built over Pinot)  wider adoption
  • 52. 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 β€˜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.
  • 53. 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 Deepak Agarwal, Igor Perisic, Arun Swami
  • 55. LinkedIn Salary (launched in Nov, 2016)
  • 56. Data Privacy Challenges Minimize the risk of inferring any one individual’s compensation data Protection against data breach No single point of failure
  • 57. Problem Statement How do we design LinkedIn Salary system taking into account the unique privacy and security challenges, while addressing the product requirements? 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)
  • 58. 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 β‡’ Copy to Hadoop (HDFS) Note: Original submission stored as encrypted objects.
  • 60. Acknowledgements Team: AI/ML: Krishnaram Kenthapadi, Stuart Ambler, Xi Chen, Yiqun Liu, Parul Jain, Liang Zhang, Ganesh Venkataraman, Tim Converse, Deepak Agarwal Application Engineering: Ahsan Chudhary, Alan Yang, Alex Navasardyan, Brandyn Bennett, Hrishikesh S, Jim Tao, Juan Pablo Lomeli Diaz, Patrick Schutz, Ricky Yan, Lu Zheng, Stephanie Chou, Joseph Florencio, Santosh Kumar Kancha, Anthony Duerr Product: Ryan Sandler, Keren Baruch Other teams (UED, Marketing, BizOps, Analytics, Testing, Voice of Members, Security, …): Julie Kuang, Phil Bunge, Prateek Janardhan, Fiona Li, Bharath Shetty, Sunil Mahadeshwar, Cory Scott, Tushar Dalvi, and team Acknowledgements David Freeman, Ashish Gupta, David Hardtke, Rong Rong, Ram
  • 61. Beyond Accuracy Performance and Cost Fairness and Bias Transparency and Explainability Privacy Security Safety Robustness
  • 63. Fairness in ML Application specific challenges Conversational AI systems: Unique bias/fairness/ethics considerations E.g., Hate speech, Complex failure modes Beyond protected categories, e.g., accent, dialect Entire ecosystem (e.g., including apps such as Alexa skills) Two-sided markets: e.g., fairness to buyers and to sellers, or to content consumers and producers Fairness in advertising (externalities) Tools for ensuring fairness (measuring & mitigating bias) in AI lifecycle Pre-processing (representative datasets; modifying features/labels) ML model training with fairness constraints Post-processing Experimentation & Post-deployment
  • 64. Explainability in ML Actionable explanations Balance between explanations & model secrecy Robustness of explanations to failure modes (Interaction between ML components) Application-specific challenges Conversational AI systems: contextual explanations Gradation of explanations Tools for explanations across AI lifecycle Pre & post-deployment for ML models Model developer vs. End user focused
  • 65. Privacy in ML Privacy-preserving model training, robust against adversarial membership inference attacks Privacy for highly sensitive data: model training & analytics using secure enclaves, homomorphic encryption, federated learning / on- device learning, or a hybrid Privacy-preserving transfer learning (broadly, privacy-preserving mechanisms for data marketplaces)
  • 66. Reflections β€œFairness and Privacy by Design” when building AI products Collaboration/consensus across key stakeholders NYT / WSJ / ProPublica / ToI / The Hindu test :)
  • 67. Thanks! Questions? S. C. Geyik, S. Ambler, K. Kenthapadi, Fairness-Aware Ranking in Search & Recommendation Systems with Application to LinkedIn Talent Search, KDD’19 [Microsoft’s AI/ML conference (MLADS’18). Distinguished Contribution Award] K. Kenthapadi, T. T. L. Tran, PriPeARL: A Framework for Privacy-Preserving Analytics and Reporting at LinkedIn, CIKM’18 K. Kenthapadi, A. Chudhary, S. Ambler, LinkedIn Salary, IEEE Symposium on Privacy-Aware Computing (PAC), 2017 [Related: our KDD’18 & CIKM’17 (Best Case Studies Paper Award) papers] Our tutorials on privacy, on fairness, and on explainability in industry at KDD/WSDM/WWW/FAccT/AAAI (combining experiences at Apple, Facebook, Google, LinkedIn, Microsoft)