Last week, I decided to pack a bag and travel to Bangkok for a week, especially after reading the NYT’ 2050s prediction of sea-level rise. I was curious about how people would react but sadly could not find any news on the local newspaper or daily conversations.
During that trip, I accidentally attended a meetup about AI privacy and security. The core topic was initialized by an interest in Thailand Personal Data Protection Act, especially towards healthcare and insurance. Despite my intensive outreach in several communities around North America, Europe, and Asia, this was the first time I heard about differential privacy and federated learning.
But let’s back to healthcare
In 2016, when starting a career as a data scientist from San Francisco, healthcare was my top choice along with education. After years, my knowledge has been accumulated by experience and effort in hackathons and collaboration with startups and academia around the world. It was how I learn an ugly truth about healthcare from seniors.
The problem is not about technology, but policies. This time I got another confirmation in the meetup from a person working in an IT section of a hospital. His concern is trying to help people understand a policy. On one side, it is designed to protect citizens from abuse of cooperation and companies using data without consensus. But it means more friction in communication and implementation for innovation.
Compared to media and e-commerce, the advertising industry, healthcare is where the pain is real and its cost is measured by human lives.
Healthcare and insurance are top industries taking privacy issues, especially consensus, as top priorities. Those are basic rights written by constitutions in many countries requiring several layers of understanding by politicians and policyholders.
But in the context of progressive advanced in technology, it is super hard for someone to grasp without knowledge of business model or computer background. We can check a testimony of CEO of Facebook — Mark Zuckerberg — before the US congress.
In Math We Trust
In the beginning, I mistakenly understand differential privacy as blockchain and federated learning as distributed computing. They are slightly different indeed.
Blockchain (Example: Bitcoin) can be broadly understood as a decentralized data management system with an emphasis on synchronization between all databases vs. differential privacy is a system designed by designed to preserve level privacy by adding noise to data.
Distributing computing (Example: Google) is a technique to parallel computing processes in different machines at the same time to accelerate process whereas federated learning is a machine learning technique with encrypted distributed data.
(Sorry if I am using too many jargons here)
Federated learning has not popular in the tech community, even from top communities in which I spend a fair amount to physically networking in North America and Europe.
Indeed, it gets quite low attention compared to deep learning though the first blog can trace back from 2017.
At the end of the meetup, I posed a question of how robustly algorithms can stand towards attacks from AIs. It was my concern in 2015 as a researcher in privacy and security.
At that moment, I read a paper about Generative Adversarial Networks — which is a game of two AIs fighting each other to decode encrypted information. Eventually, I moved to machine learning and dedicated almost at least nearly 3 relentlessly years to work on the game of 3 AIs — but it was another story.
Conclusion:
In short, federated learning and differential privacy are still unknown in the research community compared to other aspects of machine learning.
It might be a good solution for the most down-to-earth concern of policies that protects your personal data especially in the context that more governments are trying to implement AI algorithms in a race of technology.
Right now, most of the frameworks for TensorFlow and PyTorch, the two most common ai frameworks, still under development. But you can learn more from well-credited source as
A funny comic from Google AI: https://federated.withgoogle.com/
Udacity, a free course with Pytorch https://classroom.udacity.com/courses/ud185
Open mined community with 5000 members https://github.com/OpenMined/PySyft
Feel free to reach out to me at emma@blueskin.tech
Emma
