# Privacy at Facebook Scale

**Authors:** Paulo Tanaka, Sameet Sapra, Nikolay Laptev  
**arXiv:** [2006.14109](https://arxiv.org/abs/2006.14109)  
**Date:** 2020-06-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2006.14109v1/

## Key points

- Challenges faced by organizations in managing vast amounts of data collected across various aspects of their operations
- Maintaining privacy and security during audits is a critical concern
- The authors present an end-to-end system designed to detect sensitive semantic types within Facebook at scale and automatically enforce data retention and access controls
- Traditional approaches to content-based data classification are ineffective and non-scalable due to the sheer volume of data assets within Facebook
- The proposed solution combines data signals, machine learning techniques, and traditional fingerprinting methods to accurately map out and classify all data within Facebook
- The described system has been successfully implemented in production environments with impressive average F2 scores exceeding 0.9 across various privacy classes
- Represents a significant step forward in ensuring robust privacy protection and efficient data management within large-scale organizations like Facebook

## Summary

The paper "Privacy at Facebook Scale" by Paulo Tanaka, Sameet Sapra, and Nikolay Laptev addresses the challenges faced by organizations in managing vast amounts of data collected across various aspects of their operations. In today's digital age, data is constantly copied, transformed, and dispersed throughout an organization's data warehouse. As a result, maintaining privacy and security during audits becomes a critical concern. To tackle this issue, the authors present an end-to-end system designed to detect sensitive semantic types within Facebook at scale and automatically enforce data retention and access controls. Traditional approaches to content-based data classification rely on fingerprinting data and monitoring endpoints for flagged information. However, with the sheer volume of constantly evolving data assets within Facebook reaching trillions in number, this method proves to be both ineffective and non-scalable. In response to this challenge, the authors propose a novel solution that combines data signals, machine learning techniques, and traditional fingerprinting methods to accurately map out and classify all data within Facebook. This described system has been successfully implemented in production environments with impressive average F2 scores exceeding 0.9 across various privacy classes while effectively managing trillions of diverse data assets. By leveraging a combination of cutting-edge technologies and innovative strategies, this privacy system represents a significant step forward in ensuring robust privacy protection and efficient data management within large-scale organizations like Facebook.