HEIR is an open-source compiler toolchain designed to assist builders run AI fashions on encrypted knowledge with out decrypting the underlying info.
Google has launched HEIR, quick for Homomorphic Encryption Intermediate Illustration, an open-source compiler and growth platform for working AI fashions on encrypted inputs. The software goals to make homomorphic encryption extra accessible by serving to builders adapt pre-trained AI fashions that usually course of unencrypted knowledge to work with encrypted inputs.
HEIR is a part of Google’s Personal Computing Toolkit, which goals to cut back the experience required to implement homomorphic encryption. The know-how permits computations to be carried out on encrypted knowledge, enabling a service to course of delicate info with out accessing the underlying knowledge.
Google demonstrated HEIR by way of 4 private-inference functions compiled with the toolchain. These embody a content material suggestion mannequin, a credit-card fraud detector, an intrusion detection system that analyses encrypted community packets, and a hotword detector that identifies wake phrases from encrypted audio recordings. The functions had been evaluated utilizing a single-threaded CPU.
Google has collaborated with universities and {hardware} acceleration firms in growing the challenge. Its analysis companions embody Georgia Institute of Expertise, Carnegie Mellon College, UC Santa Barbara, Purdue College and the College of Edinburgh. The corporate has additionally labored with acceleration companions together with Belfort, Niobium, Cornami and Optalysys to analysis encrypted AI workloads.
Homomorphic encryption introduces computational overhead, making efficiency and price vital issues for deployment. The aim of the HEIR challenge is to cut back boundaries to experimentation by offering builders with a freely accessible toolchain and examples of how it may be utilized in several situations. The supply code for the demonstration functions is obtainable by way of GitHub.
