Hello SSN.

Arvind Ram

SSN CSE 2013 Batch

Machine Learning

Why  ML ?

To make RIGHT 

decisions

People can make RIGHT decisions!

Correct. But...

Hard to make large number of them

Decision fatigue

Leads to WRONG decisions

Birth of ML Algorithms

Decisions need DATA

Data needs Privacy

Data reveals your identity

Identity loss can have serious consequences

Ex: Gender Discrimination

Ex: Caste Discrimination

Ex: Theft Targets

We need to protect identity

From who ?

External hackers

Internal hackers

How ?

Multi level access control for extenral hackers

Internal hackers/employees

Data Anonymization

Data Anonymization

Reference: https://trustarc.com/blog/wp-content/uploads/Hash_information_tableData-anonymization-blog.png

Case Study

Netflix Case Study

Reference: https://www.youtube.com/watch?v=gI0wk1CXlsQ

Netflix Case Study

Reference: https://www.youtube.com/watch?v=gI0wk1CXlsQ

Netflix Case Study

Reference: https://www.youtube.com/watch?v=gI0wk1CXlsQ

MA Medical Records

Reference: https://www.youtube.com/watch?v=gI0wk1CXlsQ

Unique ID

Reference: https://www.youtube.com/watch?v=gI0wk1CXlsQ

Differential Privacy

 Alter data

 Won't we lose truth ?

Differential Privacy

Reference: https://towardsdatascience.com/understanding-differential-privacy-85ce191e198a

Use the data after removing noise (25%)

Plausible Deniability

$$$ LEGAL $$$

Facebook Lawsuit

Reference: https://dailyillini.com/news/2020/02/20/facebook-pay-illinois-550-million/

Facebook Lawsuit

Reference: https://dailyillini.com/news/2020/02/20/facebook-pay-illinois-550-million/

Ethics

Helps Make RIGHT decisions

Be good. Do good.

What is RIGHT ?

Depends...

On ?

On ?

Result of a decision

WRONG decisions result in bias

Microsoft Tay.ai

Microsoft Tay.ai

Reference: https://twitter.com/geraldmellor/status/712880710328139776

How to avoid this ?

• Foresee the complete potential of the solution
• Blacklist sensitive topics
• Need to eliminate such traces at training data level 
• Strong ethical testing rules

Google Vision

Google Vision Detection

Reference: https://algorithmwatch.org/en/story/google-vision-racism/

How to avoid this ?

• Know your users better
• Diversity in dataset
• Strong ethical testing rules

Amazon Recruitment

Amazon Recruitment Tool

Reference: https://fortune.com/2018/10/10/amazon-ai-recruitment-bias-women-sexist/

How to avoid this ?

• Never transfer human bias into dataset
• Clearly define and analyze decisions 
• Strong ethical testing rules

Engineer Responsibly!

Thank You