Google is making private AI practical with homomorphic encryption

Google has unveiled HEIR, an open-source compiler designed to make private AI inference practical. This tool addresses the challenge of balancing data security and functionality, particularly in sensitive sectors like healthcare and finance. Homomorphic encryption, a cutting-edge encryption method, enables computations on encrypted data, ensuring privacy without compromising functionality.

HEIR operates by converting pre-trained AI models to process encrypted inputs, thereby allowing cloud services to perform operations such as content recommendations without exposing user data. Unlike hardware-based solutions, homomorphic encryption provides strong security and privacy guarantees purely through cryptography. However, manually implementing this technology is complex and typically requires a team of cryptographers.

To address this usability issue, Google's research team developed HEIR, a compiler toolchain and development platform. HEIR streamlines the process, enabling non-experts to incorporate encrypted inference into their applications. In just four years since its initial announcement, HEIR has gained widespread adoption in the homomorphic encryption community, partnering with hardware accelerator developers like Belfort, Niobium, Cornami, and Optalysys.

The advancements in homomorphic encryption have been significant, with HEIR becoming a productive research platform for cryptographers. Collaborations with several universities have resulted in several peer-reviewed publications and numerous citations. To showcase the capabilities of homomorphic encryption, Google is sharing four private inference applications compiled with HEIR, along with their latency numbers for a single-threaded CPU.

All source code is available on GitHub, demonstrating the growing accessibility and practicality of homomorphic encryption in the AI landscape.