Protecting Privacy through Homomorphic Encryption
What is Private AI?
Private AI means running analytics and machine learning on data that is never exposed. Homomorphic encryption makes this possible: a computation runs directly on encrypted data and returns an encrypted result. The data owner does not reveal the raw data, and the operator never sees it.
This book brings together whitepapers from the HomomorphicEncryption.org standardisation community, covering the schemes, the security standard and practical guidance. It also includes privacy-preserving machine-learning applications built at the Private AI Bootcamp hosted by Microsoft Research.
Use cases in the book
- Healthcare: private translation of medical data, computing on shared healthcare data, and privacy-preserving prescription drug management.
- Mental health: detecting suicidal ideation from speech without exposing the recording.
- AI marketplaces: a trusted protocol for trading machine-learning models.
- Email security: identifying spam email on encrypted data.
- Children's privacy: privacy-preserving phone usage tracking and private movie recommendations for children.
- Data collaboration: private set intersection and compute, sharing and computing across multiple data providers, and a trusted monitoring service.
Cyber defence and national intelligence
Secure and Confidential Rule Matching for Network Traffic Analysis, written with Dimitar Jetchev of Inpher (pages 81-86).
Network traffic is monitored against detection rules, the approach used by intrusion detection systems such as Suricata, which the chapter cites. Both sides hold sensitive material: the traffic being inspected, and the rules, which often encode threat intelligence. Homomorphic encryption allows matching to take place without exposing the traffic to the analyst or the rules to the party being monitored. The chapter's references include Canada's Communications Security Establishment (CSE) Innovation Challenge.
The same pattern applies wherever organisations must share detection capability without sharing the underlying data: banks, telcos, government agencies and security partners.