Federated Learning – A case for privacy-preserving yet collaborative machine Learning

INTRODUCTION

Standard machine learning approaches for large-scale applications require centralizing the training data on one machine or a cloud server. Currently, all individual devices’ data is first uploaded on a centralized cloud server, where a mathematical function or a model is fed all the data for achieving better prediction accuracy. The model stays in the cloud server while its parameters are getting tuned. Finally the updated model is sent to the edge devices for prediction on unseen data.

The standard paradigm is a big hurdle in the way of maintaining data owner’s privacy. 

Some ML use cases ( such as from the medical domain) demand vast amounts of sensitive and personal data for better prediction accuracy. This cannot be achieved since people tend to avoid third parties getting access to their personal fitness data. The situation is further made difficult by multiple compliance and data security laws in order to protect patient’s privacy. This situation demands a new paradigm, wherein people/ medical consultants could reap the benefits of AI/ML in medical services without sharing  their personal details. Below we explain this new paradigm on which the whole future of AI/ML depends on as governments around the world have recently been passing new laws to regulate the data access by third parties. 

Federated Learning enables devices/ data centers to collaboratively learn a shared prediction model while keeping all the training data on the data owner’s device, decoupling the ability to do machine learning from the need to store the data in the cloud. 

HIGH LEVEL ARCHITECTURE

It functions like this: your device downloads the current model from a central cloud server (which is conducting & monitoring the federated learning process), improves the model/algorithm by learning from the data on your device (i.e pc, mobile, smart watch, a local data center) and then summarizes the changes as a small focused update. Only this updated model is then sent to the cloud, using encrypted communication, where it is averaged with other user updates to improve the shared model. All the training data remains on your device, and no individual updates are stored in the cloud.

Your device then imports the globally updated model (which has the flavor of the  vast amount of data it learned with) for prediction on new data generated by your device. 

ADVANTAGES

  • Data Security & Privacy Allow  Model training On Highly Sensitive Datasets
  • Real Time prediction on edge devices
  • Offline prediction
  • Minimal infrastructure

USE CASES (Medical domain)

Proactive patient care & Better utilization of hospital resources:

Researchers at MIT CSAIL, Harvard University Medical School, and Tsinghua University’s Academy of Arts and Design have recently developed a federated learning architecture for Electronic Medical Records ( EMR ) model training. In a  published paper,  they describe an architecture that sources data from 208 local hospitals around the U.S, learns a model for each community, and aggregates the models into a single effective model on the cloud server. The updated model is then sent to each local hospital server for execution. 

The researchers use the architecture to predict patient mortality and patient’s stay time at hospital. Predicting these 2 features could help in achieving proactive patient care and optimization of hospital resources.[1]

Decision support systems/ Immediate Point of care testing:

NVIDIA, King’s College London, and Owkin team up to connect hospitals across the UK with federated learning to provide immediate point of care diagnoses. In contrast to traditional medical diagnoses, in which scans are first sent for further analysis by specialists, point of care testing allows immediate diagnosis from X-rays, CT scans or MRI at the time of the patient-doctor interaction. This specialist-independent diagnoses could only be made possible through model learning from diverse patterns hidden in a vast data sourced from multiple hospitals.   

King’s College London will apply the federated learning architecture to classify stroke and neurological impairments, determine the underlying causes of cancers, etc. [2]

Quick identification and development of new drugs:

Melloddy, an AI firm in collaboration with 17 partners: 10 leading pharmaceutical companies, including Amgen, Bayer, GSK, Janssen Pharmaceutica and Novartis & top European universities KU Leuven and the Budapest University of Technology and Economics among others, will use federated learning to be able to collaborate and build the world’s largest drug compound dataset for AI training without sacrificing data privacy. [3]

LIMITATIONS/ CHALLENGES

  • Developing an infrastructure which can keep pace with the dynamic and continuous learning involved
  • Maintaining massively distributed systems
  • Unreliable connectivity
  • Unbalanced data in terms of bias or feedback.
  • Running optimization algorithms across highly distributed data sets

FRAMEWORKS AVAILABLE

  • Federated Learning by OpenMined
  • Tensorflow Federated
  • FATE (Federated AI Technology Enabler) 
  • IBM Federated Learning
  • Federated Learning by NVIDIA Clara

REFERENCES

1.https://arxiv.org/ftp/arxiv/papers/1903/1903.09296.pdf

2.https://www.nvidia.com/content/dam/en-us/Solutions/data-center/gated-resources/hc-casestudy-kings-college-london.pdf

3.https://blogs.nvidia.com/blog/2019/08/08/pharma-melloddy-ai-drug-discovery-consortium/

Find out how Federated Learning can accelerate results across your business — from employees and customers to products and operations.

Get in touch and schedule an appointment today with one of our consultants: