HEIR is an open-source compiler toolchain designed to help developers run AI models on encrypted data without decrypting the underlying information.
Google has released HEIR, short for Homomorphic Encryption Intermediate Representation, an open-source compiler and development platform for running AI models on encrypted inputs. The tool aims to make homomorphic encryption more accessible by helping developers adapt pre-trained AI models that normally process unencrypted data to work with encrypted inputs.
HEIR is part of Google’s Private Computing Toolkit, which aims to reduce the expertise required to implement homomorphic encryption. The technology allows computations to be performed on encrypted data, enabling a service to process sensitive information without accessing the underlying data.
Google demonstrated HEIR through four private-inference applications compiled with the toolchain. These include a content recommendation model, a credit-card fraud detector, an intrusion detection system that analyses encrypted network packets, and a hotword detector that identifies wake words from encrypted audio recordings. The applications were evaluated using a single-threaded CPU.
Google has collaborated with universities and hardware acceleration companies in developing the project. Its research partners include Georgia Institute of Technology, Carnegie Mellon University, UC Santa Barbara, Purdue University and the University of Edinburgh. The company has also worked with acceleration partners including Belfort, Niobium, Cornami and Optalysys to research encrypted AI workloads.
Homomorphic encryption introduces computational overhead, making performance and cost important considerations for deployment. The purpose of the HEIR project is to reduce barriers to experimentation by providing developers with a freely accessible toolchain and examples of how it can be applied in different scenarios. The source code for the demonstration applications is available through GitHub.
